Curated catalog / 精选目录

Reviewed 2026-11-12, catalog schema 2. 119 entries, with 51 installable/configurable integrations and 68 reference-only tools. The installable set contains 24 skills, 6 rule packs, 19 MCP servers, and 2 MCP-to-LSP bridges. The 18 LSP entries comprise the two installable bridges (Serena, lsp-mcp-server), the three reference-only bridges (agent-lsp, mcpls, lspyx), and 13 standalone language-server references.

Stars are dated repository-level snapshots, not scores for individual skills or rules. Multiple entries from ECC, Superpowers, agent-skills, or another monorepo share the same repository evidence. The executable source of truth is catalog.json; every entry records its source, license, review date, and installation status.

Browse and select

apd profiles
apd list --installable
apd list --kind rules
apd list --kind lsp
apd list --category framework-best-practices --installable
apd init --agents codex --profile django --dry-run

install: null means the CLI lists the tool and its upstream setup guidance. An installable skill copies the pinned skill directory. A rule pack selects and activates reviewed Markdown rules for the chosen agent. A configurable MCP entry writes a pinned server launch configuration; the coding agent starts it later. Serena uses that MCP path to expose language-server capabilities.

Presets

Profiles select relevant tools instead of loading every available rule and framework. requires adds shared rules transitively. For example, FastAPI rules add Python rules, which add common rules.

Preset	Selected entries	Dependencies added
`starter`	`systematic-debugging`, `test-driven-development`, `verification-before-completion`, `context7`, `ecc-common-rules`	None
`python`	`systematic-debugging`, `test-driven-development`, `verification-before-completion`, `context7`, `python-testing-patterns`, `python-packaging`, `fetch`, `ecc-python-rules`, `python-patterns`	`ecc-common-rules`
`web`	`systematic-debugging`, `test-driven-development`, `verification-before-completion`, `context7`, `playwright`, `ecc-typescript-rules`, `frontend-patterns`	`ecc-common-rules`
`engineering`	`ecc-common-rules`, `coding-standards`, `api-design`, `database-migrations`, `deployment-patterns`, `systematic-debugging`, `test-driven-development`, `verification-before-completion`	None
`security`	`ecc-common-rules`, `security-review`	None
`django`	`ecc-python-rules`, `python-patterns`, `django-patterns`, `django-security`, `python-testing-patterns`, `context7`	`ecc-common-rules`
`fastapi`	`ecc-fastapi-rules`, `python-patterns`, `fastapi-patterns`, `api-design`, `security-review`, `python-testing-patterns`, `context7`	`ecc-common-rules`, `ecc-python-rules`
`springboot`	`ecc-java-rules`, `springboot-patterns`, `springboot-security`, `api-design`, `database-migrations`, `context7`	`ecc-common-rules`
`typescript`	`ecc-typescript-rules`, `coding-standards`, `frontend-patterns`, `backend-patterns`, `nextjs-turbopack`, `security-review`, `context7`	`ecc-common-rules`
`golang`	`ecc-golang-rules`, `golang-patterns`, `api-design`, `security-review`, `context7`	`ecc-common-rules`
`lsp`	`serena`	None

Engineering skills, standards, and delivery

ID / source	Use case	Delivery	Stars	License
`systematic-debugging`	Root-cause investigation, reproducible failures, and hypothesis-driven debugging.	Install skill	306,063	MIT
`test-driven-development`	A test-first workflow with guidance for meaningful regression tests.	Install skill	306,063	MIT
`verification-before-completion`	Require fresh verification evidence before claiming a change is complete.	Install skill	306,063	MIT
`coding-standards`	Shared naming, readability, error-handling, immutability, and code-quality conventions.	Install skill	290,505	MIT
`api-design`	REST resources, HTTP semantics, pagination, versioning, error contracts, and rate limits.	Install skill	290,505	MIT
`database-migrations`	Expand-contract schema changes, safe indexes, batched backfills, and migration-tool workflows.	Install skill	290,505	MIT
`deployment-patterns`	Deployment strategies, container builds, health checks, CI/CD, rollback, and production readiness.	Install skill	290,505	MIT
`code-review-and-quality`	Structured review of a diff or pull request across correctness, readability, architecture, security, and performance before merge.	Install skill	111,131	MIT
`observability-and-instrumentation`	Logging, metrics, tracing, and alerting decisions that make production behavior diagnosable.	Install skill	111,131	MIT
`open-code-review`	Run an AI review over a diff with deterministic analysis pipelines plus a tuned review model, with matching agent skills and plugins.	Reference only	49,294	Apache-2.0
`graphify`	Build a local, deterministic knowledge graph of a repository, its docs, SQL schemas and configuration, and answer architecture, dependency and change-impact questions through a coding-agent skill with no vector store.	Reference only	142,432	Apache-2.0
`brooks-lint`	Six review modes for architecture decay, technical debt, test weakness, and remediation, using book-derived risk codes and a 0-100 health score, packaged for fourteen coding agents.	Reference only	1,561	MIT
`planning-with-files`	Keep the plan in files instead of context: `.planning/` plan, progress and findings documents, per-turn re-injection against context rot, session recovery after `/clear` or compaction, and a deterministic completion gate.	Install skill	27,973	MIT
`retentioneering`	Product analytics from user-level event logs: paths, transitions, funnels, segments, retention, step matrices and sequence patterns, with a bundled API map, analysis recipes and a log-inspection script.	Install skill	946	Apache-2.0

Python, languages, and framework practices

ID / source	Use case	Delivery	Stars	License
`python-testing-patterns`	Practical pytest fixtures, parametrization, mocking, async testing, and coverage guidance.	Install skill	41,868	MIT
`python-packaging`	Build Python libraries and CLIs with package metadata, wheels, and release workflows.	Install skill	41,868	MIT
`python-dev-skills`	Claude Code skills for the Python library lifecycle — packaging, testing, API design, security audit, performance, release management, MCP servers, FastAPI/Postgres web apps — plus shared and other-language bundles.	Reference only	111	MIT
`python-patterns`	Python idioms, type hints, error handling, resource management, and maintainable code structure.	Install skill	290,505	MIT
`fastapi-patterns`	FastAPI application structure, Pydantic v2, async dependencies, service layers, and HTTP testing.	Install skill	290,505	MIT
`django-patterns`	Django and DRF architecture, ORM access, caching, middleware, and service organization.	Install skill	290,505	MIT
`frontend-patterns`	React and Next.js component composition, state, data fetching, forms, performance, and accessibility.	Install skill	290,505	MIT
`nextjs-turbopack`	Next.js 16+ bundler selection, development performance, and middleware-to-proxy migration guidance.	Install skill	290,505	MIT
`backend-patterns`	Node.js, Express, and Next.js API architecture, data access, caching, and observability.	Install skill	290,505	MIT
`springboot-patterns`	Spring Boot REST layers, transactional services, repositories, caching, and production defaults.	Install skill	290,505	MIT
`golang-patterns`	Idiomatic Go errors, interfaces, concurrency, package organization, and resource lifecycle patterns.	Install skill	290,505	MIT
`ui-ux-pro-max`	Query a local design database for a UI task — 80 styles, 192 palettes, 74 font pairings, 120 UX guidelines and per-stack reasoning rules for 22 stacks — and produce a project design system instead of a generic palette.	Reference only	147,361	MIT

Security review and scanning

ID / source	Use case	Delivery	Stars	License
`security-review`	Review authentication, authorization, untrusted input, secrets, and sensitive application changes.	Install skill	290,505	MIT
`django-security`	Django authentication, authorization, request protections, and production security settings.	Install skill	290,505	MIT
`springboot-security`	Spring Security authentication, authorization, validation, secrets, headers, and release checks.	Install skill	290,505	MIT
`semgrep`	Rule-based static analysis for security issues and code patterns across multiple languages.	Reference only	17,306	LGPL-2.1 (community engine); additional offerings/rules have separate terms
`bandit`	Find common Python security issues through AST-based static analysis.	Reference only	8,352	Apache-2.0
`pip-audit`	Audit Python environments and dependency specifications against known vulnerability data.	Reference only	1,386	Apache-2.0
`osv-scanner`	Scan lockfiles, manifests, SBOMs, and containers across ecosystems against the OSV vulnerability database.	Reference only	11,342	Apache-2.0
`gitleaks`	Detect committed or working-tree secrets using configurable rules and allowlists.	Reference only	30,384	MIT
`skillspector`	Scan a skill directory, MCP server, Git URL, archive, or MCP registry snapshot for prompt injection, exfiltration, credential access, obfuscation, and supply-chain risk before installing it.	Reference only	24,520	Apache-2.0
`pipelock`	Mediate an agent's egress and MCP traffic against a policy file, scanning requests and responses and signing a receipt for every decision.	Reference only	987	Apache-2.0
`trailofbits-skills`	Trail of Bits plugin marketplace of 47 Agent Skills for security review — audit context building, differential and variant analysis, Semgrep rule authoring, patch validation, supply-chain auditing — plus development workflows including `modern-python`, property-based testing and static analysis.	Reference only	8,109	CC-BY-SA-4.0
`mantis`	Run agentic vulnerability research as a method: history review, semantic index, threat model, hypothesis planning, deduplication, static verification, reproduction, chaining, adversarial patching and severity calibration, plus an ADK harness and a pull-request action.	Reference only	2,759	Apache-2.0

Portable agent rules

ID / source	Use case	Delivery	Stars	License
`ecc-common-rules`	Portable cross-language engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ecc-python-rules`	Portable python-specific engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ecc-typescript-rules`	Portable typescript-specific engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ecc-golang-rules`	Portable golang-specific engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ecc-java-rules`	Portable java-specific engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ecc-fastapi-rules`	Portable fastapi-specific engineering guidance selected from ECC rules.	Install rules	290,505	MIT
`ruler`	Apply one `.ruler/` source of instructions and MCP settings to the native rule, skill, subagent and MCP files of more than thirty coding agents, keeping generated files out of version control.	Reference only	2,959	MIT

Documentation, browser, and MCP ecosystem

ID / source	Use case	Delivery	Stars	License
`context7`	Retrieve library documentation and version-specific examples for coding tasks.	Configure MCP	63,759	MIT
`playwright`	Browser inspection and interactive web testing through accessibility snapshots.	Configure MCP	38,702	Apache-2.0
`fetch`	Retrieve web pages as text/Markdown for documentation and research.	Configure MCP	91,599	Apache-2.0 / MIT; docs CC-BY-4.0
`code-review-graph`	Serve a Tree-sitter code graph over MCP for review context, change impact, and affected-test hints.	Configure MCP	32,459	MIT
`mcp-tools-py`	Run pylint, pytest, mypy, Ruff, Bandit, formatting, and symbol refactors on a Python project through one scoped MCP server.	Configure MCP	20	MIT
`fallow`	Report TypeScript and JavaScript complexity, duplication, architecture, styling, and dead-code findings over MCP.	Configure MCP	5,273	MIT
`repomix`	Pack a repository into one AI-friendly file and serve packing, search, and reading of the output over MCP.	Configure MCP	29,100	MIT
`roam-code`	Serve a local codebase index over MCP for symbol search, dependency and coupling maps, dead code, taint, and change-impact questions.	Configure MCP	516	Apache-2.0
`headroom`	Compress tool output, logs, files, and RAG chunks locally before they reach the model, and retrieve the originals over MCP.	Configure MCP	78,668	Apache-2.0
`hippo-memory`	Long-term memory for coding agents over MCP from a local store that strengthens recalled memories and lets mistakes decay.	Configure MCP	791	MIT
`reticle`	Verify the project's running web app from inside a real browser and get pass/fail/cannot-tell verdicts with the `file:line` behind them.	Configure MCP	1,308	Apache-2.0
`seleniumbase`	Drive a real Chrome through SeleniumBase's pure CDP mode over MCP: open pages, inspect and click elements, type, run JavaScript, manage cookies and storage, assert, and interact with CAPTCHA challenges.	Configure MCP	13,228	MIT
`dbhub`	Execute SQL and search object schemas on PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, and SQLite through a two-tool, token-efficient MCP server.	Configure MCP	3,801	MIT
`squirrelscan`	Audit a deployed site over MCP and the CLI and return SEO, accessibility, performance, content and link findings an agent can fix.	Configure MCP	308	MIT
`mcp-toolbox`	Connect an agent to PostgreSQL, MySQL, SQLite, SQL Server, BigQuery, and other databases through prebuilt or hand-defined SQL tools.	Reference only	16,792	Apache-2.0
`github-mcp`	Official GitHub tools for repositories, issues, and pull requests.	Reference only	33,999	MIT
`chrome-devtools`	Inspect browser performance, network activity, and runtime behavior.	Reference only	54,261	Apache-2.0
`sequential-thinking`	A reference server for structured problem-solving tool calls.	Reference only	91,599	Apache-2.0 / MIT; docs CC-BY-4.0
`codebase-memory-mcp`	Local knowledge graph over code, documentation, configuration, and infrastructure for navigation and change impact.	Reference only	47,734	MIT
`chunkhound`	Index a codebase locally and answer code questions over MCP with citations from current code, git history, and docs.	Reference only	1,455	MIT
`ripwire`	Ranked, deterministic call graph that maps a repository before an agent reads it and checks an edit's blast radius, reachable tests, and quality regressions.	Reference only	2,513	Apache-2.0
`codeboarding`	Render a codebase as interactive architecture maps plus Markdown, HTML, MDX or reStructuredText documentation for review and a coding agent's next prompt.	Reference only	2,663	MIT
`mcp-context-forge`	Federate MCP, A2A and REST/gRPC servers behind one governed endpoint with a registry, virtual servers, auth, rate limiting, and an admin UI.	Reference only	4,752	Apache-2.0
`powercontext`	Server-side memory, handoff and experience layer for coding agents, exposed as an HTTP contract with a Python client, CLI, dashboard and a curated MCP projection over revisioned artifacts.	Reference only	1,575	Apache-2.0
`aoci-code`	Keep a Git-versioned, agent-written map of a repository's code, configuration, tests and database schema through a local Go MCP server, so an agent reads the index instead of re-reading the tree.	Reference only	1,897	FSL-1.1-MIT
`semble`	Answer code questions from a CPU-only repository index instead of grepping and reading files, over MCP (`search`, `find_related`), a CLI, an agent skill or a sub-agent.	Configure MCP	6,403	MIT

Language intelligence and LSP

ID / source	Use case	Delivery	Stars	License
`serena`	Expose semantic code navigation, references, diagnostics, and edits to coding agents through an MCP-to-LSP bridge.	Configure LSP bridge	30,759	GPL-3.0-or-later (application); MIT (SolidLSP)
`lsp-mcp-server`	Bridge MCP clients to whichever language server a project configures, exposing navigation, references, diagnostics, completion, and edits as 29 tools.	Configure LSP bridge	24	MIT
`agent-lsp`	Orchestrate native language servers into batch, refactor, and speculative-edit workflows for coding agents.	Reference only	166	MIT
`mcpls`	Bridge MCP to rust-analyzer, pyright, gopls, and other language servers and expose 31 code tools to agents.	Reference only	80	MIT OR Apache-2.0
`lspyx`	Answer Python definition, reference, outline, diagnostic and rename-preview queries over MCP and the CLI through `ty` and Ruff instead of a language server.	Reference only	0	MIT
`ruff`	Python linting, formatting, and the built-in Ruff language server for diagnostics and code actions.	Reference only	50,317	MIT
`pyright`	Python type checking and language-server support for navigation and diagnostics.	Reference only	15,707	MIT
`ty`	Rust-based Python type checker and language server with editor integrations; upstream labels it beta.	Reference only	19,973	MIT
`pyrefly`	Rust-based Python type checker and language server with navigation, completion, inlay hints, and semantic highlighting.	Reference only	7,172	MIT
`python-lsp-server`	Jedi-based Python completion and navigation with configurable linting and formatting plugins.	Reference only	2,605	MIT
`typescript-language-server`	Language Server Protocol wrapper around TypeScript language services for JavaScript and TypeScript.	Reference only	2,587	Apache-2.0; MIT portions
`gopls`	Official Go language server providing code navigation, diagnostics, and refactoring.	Reference only	8,009	BSD-3-Clause
`rust-analyzer`	Rust language server with code navigation, completion, diagnostics, and assists.	Reference only	16,949	MIT OR Apache-2.0
`clangd`	C/C++ language server with completion, cross-references, diagnostics, and refactoring support.	Reference only	2,315	Apache-2.0 WITH LLVM-exception
`tombi`	Schema-driven TOML formatter, linter, and language server for project configuration such as `pyproject.toml`.	Reference only	1,192	MIT
`agnix`	Lint and auto-fix agent configuration — `CLAUDE.md`, `AGENTS.md`, `SKILL.md`, hooks, plugin and MCP files — with 758 per-client rules, shipped as a CLI with an `agnix-lsp` language server and an `agnix-mcp` server.	Reference only	471	MIT OR Apache-2.0
`basedpyright`	Pyright fork with Pylance features, configurable checking modes, a baseline file and an `--outputjson` report, shipping the `basedpyright-langserver` server.	Reference only	3,647	MIT
`pytest-language-server`	Rust language server for pytest: fixture definitions across `conftest.py` and plugins, `yield` navigation, fixture call hierarchy, completion, references and diagnostics.	Reference only	124	MIT

Specification-driven development

ID / source	Use case	Delivery	Stars	License
`openspec`	Spec-driven change workflows with per-agent skills and initialization.	Reference only	73,139	MIT
`spec-kit`	Specification, planning, and implementation templates for coding agents.	Reference only	143,181	MIT
`spec-kitty`	Keep specifications, work packages, acceptance criteria, review state, and merge decisions in the repository and run each work package in its own git worktree.	Reference only	1,716	MIT
`elephant-goldfish`	Draft a software design and then attack it with a ten-agent review ensemble before implementation, and report drift after the change.	Reference only	15,063	Apache-2.0

Skill ecosystems and discovery

ID / source	Use case	Delivery	Stars	License
`anthropic-skills`	Official examples of reusable agent skills and document workflows.	Reference only	184,157	Per-skill licenses; see upstream
`vercel-agent-skills`	Frontend and deployment-oriented skills maintained by Vercel.	Reference only	33,096	See upstream and individual skills
`skills-cli`	Discover and install Agent Skills across supported coding agents.	Reference only	36,801	MIT
`awesome-copilot`	Community-contributed GitHub Copilot instructions, custom agents, skills, and prompts for Python and web projects.	Reference only	40,570	MIT
`skillevaluator`	Score agent skill quality, detect overlap between skills, and measure a skill's effect on live agent runs.	Reference only	612	Apache-2.0
`ai-rulez`	Compile one rules, skills, and MCP source into native instruction files for dozens of coding agents, with lock files, signing, approvals, and policy checks.	Reference only	166	MIT
`cross-code-organizer`	Inventory what each coding agent will actually load — configuration, memories, skills, and MCP servers — edit those files in place, and scan MCP tool descriptions for poisoning.	Reference only	386	MIT
`skillsaw`	Lint, autofix and baseline the skills, plugin manifests, hooks and instruction files a coding agent loads across the major agent ecosystems.	Reference only	71	Apache-2.0
`agents-cli`	Seven Agent Skills plus a CLI for building, evaluating, deploying, observing and publishing ADK agents on Google Cloud.	Reference only	6,315	Apache-2.0
`agent-skills-spec`	The Agent Skills specification plus `skills-ref`, the Python reference validator that checks a skill directory and renders the properties and prompt block a client would load.	Reference only	27,113	Apache-2.0
`claude-plugins-official`	Anthropic's vetted Claude Code plugin marketplace: 317 skills, agents, hooks, commands and MCP servers plus 14 partner plugins, installed with `/plugin install <name>@claude-plugins-official`.	Reference only	46,870	Apache-2.0

Python project toolchain

ID / source	Use case	Delivery	Stars	License
`uv`	Python environments, dependency management, and isolated tool execution.	Reference only	91,242	Apache-2.0 OR MIT
`pytest`	Python testing with fixtures and a broad plugin ecosystem.	Reference only	14,649	MIT
`hypothesis`	Property-based testing for Python that generates inputs and shrinks failing examples.	Reference only	9,113	MPL-2.0
`pre-commit`	Run configured repository checks before commits.	Reference only	15,664	MIT
`deptry`	Detect missing, unused, and misplaced dependencies in a Python project.	Reference only	1,498	MIT
`orm-lens`	Answer Django model questions over MCP from static parsing — relations, ER diagrams, migration risk, N+1 candidates, and schema drift — with SQLAlchemy preview support.	Configure MCP	96	MIT
`misata`	Declare the facts a dataset must hold and get multi-table synthetic test rows over MCP, with a certificate a DuckDB run re-checks, next to a CLI and a pytest plugin.	Configure MCP	75	MIT
`mustel`	Run Ruff, Bandit, Oxlint and pip-audit over MCP for the project it is launched in, with a compressed AST code map and an audit mode for dependency vulnerabilities.	Configure MCP	19	MIT
`ast-grep`	Search, lint and rewrite code by AST pattern instead of text, over tree-sitter grammars including Python, installed as the `ast-grep-cli` package in a project environment.	Reference only	16,554	MIT
`archy`	Assert Python architecture on the import graph — blast radius, affected tests, cycles, coupling and declared layers — as a CLI that exits non-zero and an MCP server a session can call before committing.	Configure MCP	8	MIT

Building MCP integrations

ID / source	Use case	Delivery	Stars	License
`fastmcp`	Build Python MCP servers and clients with a high-level API.	Reference only	28,217	Apache-2.0
`mcp-python-sdk`	Official Python SDK for implementing MCP servers and clients.	Reference only	24,684	MIT
`mcp-inspector`	Launch, inspect, and call any MCP server from a web UI, TUI, or CLI while developing or debugging an integration.	Reference only	11,216	MIT and Apache-2.0 (relicensing in progress)
`arcade-mcp`	Build Python MCP servers with decorator tools, resources, prompts, sampling and elicitation, framework-owned OAuth, and an `arcade-mcp` CLI for scaffolding, running and deploying servers.	Reference only	1,088	MIT

Source pins and installation boundaries

Skill and rule archives are pinned to full commits and verified against a recorded SHA-256. Skills retain their complete selected directories and supporting documents. Rules use explicit file allowlists, keep upstream originals and attribution, and generate an adapted activation form for each supported agent. See ECC migration for the exact selection and transformations.

Source	Pinned revision	Archive SHA-256
obra/superpowers	`2ce8aa15d05bd57213e6f746f953aafb7ded7a8b`	`da34247c51339cdd6175b3590261f16e25b08d46669afd5b6fc3f842ab303e20`
wshobson/agents	`0874a27b17a22e2b7c2e2bcd725768e0803cad06`	`468c5d32d4a24ea9629c7713d209bbb46cf462ac487d3bc442414aec4ed54c43`
affaan-m/ECC	`25278e7d7b6cf6e6a4e7a6c8e10f14ff22f0eb5b`	`7979cc21e2093aed6218fa1709373eff5eb8641b0c0a5128c68db91bbedd7bb1`
addyosmani/agent-skills	`aeac41d4d005999432bb7312f7ff06383a91d59d`	`0211a580aab7d45c73886d6f96d2f0679416119b4f9add08b6ceea5c0c15fbf1`
OthmanAdi/planning-with-files	`2362ca385ad74f061aa0e7b73f56543a4b7e6df2`	`cf60990c520c8f5d6fff876524450ce67172ad8ce785213a9bc50dbb06ac2227`
retentioneering/retentioneering-tools	`68f699f67674375c59ea26c6a719e91e83905f51`	`c0fe070c7ad635c57b69f4d24ce394b6f21664c234ccdcf24db49af554568781`

Service	Pinned launcher	Runtime / setup
`context7`	`npx -y @upstash/context7-mcp@4.4.4`	Node.js >=20.18.1
`playwright`	`npx -y @playwright/mcp@0.0.83`	Node.js >=18
`fetch`	`uvx mcp-server-fetch==2026.10.24`	Python >=3.10
`serena`	See exact Serena command; v2.1.0 at `5e4f5cd173c7ff3eed37ee3f1de6c266b937c5dc`	uv/uvx + Git; Python 3.13 selected by uvx; language-specific dependencies
`code-review-graph`	`uvx --from code-review-graph==2.3.9 code-review-graph serve --repo {project}`	uv/uvx; Python >=3.10 selected by uvx
`mcp-tools-py`	`uvx --from mcp-tools-py==0.1.10 --with mcp<2 mcp-tools-py --project-dir {project}`	uv/uvx; Python >=3.11 selected by uvx; pylint/pytest/mypy for the checks
`fallow`	`npx --yes --package fallow@3.37.0 fallow-mcp`	Node.js >=22; the server analyses the directory it starts in
`repomix`	`npx -y repomix@1.18.1 --mcp`	Node.js >=22; without `--sandbox` the server can read any path the host user can and trusts the config of the directory it packs
`roam-code`	`uvx --python 3.12 --from 'roam-code[mcp]==14.1.0' roam mcp`	uv/uvx; Python 3.12 pinned; indexes the directory it starts in, state in `.roam/`
`orm-lens`	`uvx --python 3.12 --from 'orm-lens[mcp]==1.16.0' orm-lens-mcp`	uv/uvx; Python 3.12 pinned; static parsing only, no database or settings module
`headroom`	`uvx --python 3.13 --from 'headroom-ai[mcp]==0.42.0' headroom mcp serve`	uv/uvx; Python 3.13 pinned; MCP server mode only, no model download or proxy
`lsp-mcp-server`	`npx -y lsp-mcp-server@1.1.27`	Node.js >=18; the target language server must be installed and on PATH
`hippo-memory`	`npx -y hippo-memory@3.2.2 mcp`	Node.js >=22.16; the memory store is local (`~/.hippo`, or `.hippo/` in a project after `hippo init`)
`reticle`	`npx -y @reticlehq/server@3.9.0 mcp`	Node.js >=20.11; a supported browser and the project's own running app are prerequisites for verification
`misata`	`uvx --python 3.12 --from 'misata[mcp]==0.9.8.1' misata-mcp`	uv/uvx; Python 3.12 pinned; only the MCP server is configured, not the CLI, pytest plugin, or studio
`seleniumbase`	`uvx --python 3.13 --from 'seleniumbase[mcp]==4.57.1' seleniumbase-mcp`	uv/uvx; Python 3.13 pinned (the `mcp` extra needs >=3.10); a Chrome/Chromium installation is a prerequisite for the browser tools
`dbhub`	`npx -y @bytebase/dbhub@1.5.1 --transport stdio --demo`	Node.js >=22.5 for the built-in `node:sqlite`; `--demo` serves the bundled sample SQLite database and is replaced by `--dsn` for a real database
`mustel`	`uvx --python 3.12 --with mcp<2 --from mustel==0.3.4 mustel serve`	uv/uvx; Python 3.12 pinned (the package needs >=3.10 while this host's system Python is 3.9); published `0.3.4` leaves `mcp` uncapped, so the recipe adds `--with mcp<2`; the server analyses the directory it starts in, and its `bootstrap` tool is never called
`squirrelscan`	`npx -y squirrelscan@0.0.173 mcp`	Node.js >=18; the audit fetches the target site over HTTP, so the server needs network access and an explicit site per call; the hosted endpoint and the signed-in report scopes need the user's own account
`semble`	`uvx --python 3.12 --from 'semble[mcp]==0.7.3' semble`	uv/uvx; Python 3.12 pinned; the `mcp` extra caps `mcp<3.0`; the server indexes the directory it starts in, and the first run downloads the embedding model and grammars
`archy`	`uvx --python 3.12 --from archy==0.55.1 archy mcp`	uv/uvx; Python 3.12 pinned; static import analysis of the directory it starts in, no state written and no test run

Context7: API keys are optional in the pinned stdio package guidance for higher limits/private repositories. Service access and policy can change independently of the package version. Configure credentials through the chosen agent; do not commit secrets.
Playwright: A supported browser is needed for browser operations. Configuring the server does not install browser binaries.
Fetch: Requires uv/uvx and a compatible Python. No private-IP bypass is enabled by this recipe.
Serena: The project path is resolved into the MCP configuration. First startup may download language-server dependencies and create Serena-owned state. The installer does not manage or uninstall those upstream caches. See language intelligence and setup.
lsp-mcp-server: The second installable bridge and the only TypeScript one here: it drives whichever language server a project configures, so Python coverage depends on pyright, Python LSP Server, or another LSP binary installed separately. It pins npm release 1.1.27 (upstream tag v1.1.27, commit 2b9005387fbf5ebd101ec84cfc34a8196bf9d347), reads per-project server settings from .lsp-mcp.json, and upstream recommends a longer client tool timeout because the first call starts a language server and its initial analysis. The npm package also ships a SKILL.md that teaches tool selection and says it should be re-copied when the server version changes; only the MCP launch configuration is written here. The smoke check completed initialize and listed 29 tools, but no language server was started through it and no project file was read by a tool call, so the LSP round trip itself is unverified.
Native language servers: Pyright, basedpyright, Pyrefly, Python LSP Server, Ruff, ty, pytest-language-server, TypeScript Language Server, gopls, rust-analyzer, clangd, and Tombi remain reference-only. Their native LSP commands need an LSP client; they are not interchangeable with MCP servers.
agent-lsp: One of the two reference-only LSP bridges next to the installable Serena and lsp-mcp-server recipes. It ships as a Go binary from an upstream install script, release archive, or container image, and requires the underlying language servers separately; there is no registry package or dependency version for this catalog to pin, so apd writes only the Serena and lsp-mcp-server recipes.
mcpls: The other reference-only bridge. It is distributed as a Rust binary through cargo install mcpls, an upstream install script, release archives, or a container image, and needs at least one language server (for example pyright) installed separately. Enabling it next to Serena or lsp-mcp-server would give the same project two bridges over the same servers.
code-review-graph: The project path is resolved into the MCP configuration and the server indexes it with Tree-sitter; graph data stays in a .code-review-graph directory inside the project. Building or refreshing the graph is an explicit CLI or tool step, and upstream install also writes hooks, skills, and rules files that this recipe does not touch.
mcp-tools-py: A Python-only companion to the reference-only pytest, Ruff, and Bandit entries. Its --with mcp<2 pin is deliberate: published 0.1.10 imports mcp.server.fastmcp without capping mcp, so the default mcp 2.x resolution fails at startup. The checks run through the server's own interpreter unless --python-executable/--venv-path is configured, so pylint, pytest, and mypy must exist in that environment.
mustel: Python static-analysis orchestration over MCP with a dependency-audit lane: Ruff, Oxlint and local YAML patterns run locally in dev mode with a stat-based (mtime + size) cache, CI/GITHUB_ACTIONS/PRE_COMMIT switch on the Bandit and pip-audit passes, and get_code_map returns a compressed AST/regex skeleton of classes, functions and docstrings (upstream describes a 1,500-token budget) with .ipynb code cells parsed natively. The recipe pins --python 3.12 because the package requires Python >=3.10 while this host's system Python is 3.9, and adds --with mcp<2 because published 0.3.4 imports mcp.server.fastmcp without capping mcp — the same failure mode recorded for mcp-tools-py; the pinned release matches the version in main's pyproject.toml but predates the repository's October commits. It overlaps mcp-tools-py (pylint, pytest, mypy, Ruff, symbol renames) as a Python checker behind MCP, so pick per tool set: neither exposes the other's checks, and only this one carries dependency auditing. Its bootstrap tool and mustel bootstrap command rewrite ~/.cursor/mcp.json, ~/.claude.json, Claude Desktop configuration, project .cursorrules/.windsurfrules and git pre-commit hooks machine-wide; this catalog writes only the stdio launch configuration and never calls it. Ruff, Bandit, pip-audit, Semgrep and OSV-Scanner stay reference-only, so no scanner is installed by this recipe. The smoke check completed initialize (server mustel, reported version 1.30.0, protocol 2025-11-25) and listed 5 tools; no tool was called, so no project was analysed, no code map was generated and no guardrail fired.
fallow: TypeScript and JavaScript health analysis over MCP — complexity, duplication, architecture boundaries, styling drift, and dead code. The server analyses the directory it is launched in and takes no project argument. Upstream fallow agent install additionally writes an AGENTS.md task map, installs skills, registers itself, and adds git commit and push gates; this recipe writes only the launch configuration. The npm package pins a per-platform binary and needs Node.js 22 or newer.
repomix: Packs a repository into a single AI-oriented file and exposes the pack plus read/grep tools over MCP. The npm package pins a per-platform payload and needs Node.js 22 or newer. Upstream documents repomix --mcp and the npx form used here; the smoke check completed initialize and listed six tools. Without --sandbox the server can read any path the host user can and trusts repomix.config.* in whatever directory it packs, which upstream describes as code execution; --mcp --sandbox path/to/project confines the file tools to one workspace and drops config loading, remote packing, and skill generation, at the cost of remote targets.
roam-code: Local codebase intelligence over MCP — symbol search, dependency and coupling maps, dead code, taint paths, file metrics, and change-impact answers for the directory the server starts in. The launcher pins --python 3.12 because the package requires Python >=3.10 while an unpinned uvx resolution on the smoke host picked Python 3.9 and exited before the handshake. The index and its lock/state sidecars live in .roam/ (ROAM_DB_DIR relocates the store, ROAM_NO_AUTO_INDEX=1 refuses implicit builds); upstream documents roam init and roam index as explicit steps. The CLI and MCP server are Apache-2.0 and need no account or API key for local analysis, and upstream states local analysis does not upload source, indexes, findings, or telemetry; separate review and cloud services exist and are not configured here. The smoke check completed initialize and listed 17 tools, and no tool was called, so analysis quality is unverified.
headroom: A local context-compression layer: it shrinks tool output, logs, files, and RAG chunks before they reach the model, keeps the originals retrievable, and exposes headroom_compress, headroom_retrieve, and headroom_stats over MCP. The recipe pins the published PyPI release 0.42.0 with its mcp extra and --python 3.13, because the package requires Python 3.10 or newer and an unpinned uvx resolves the interpreter itself. The CLI ships only in that package — the npm headroom-ai package is the TypeScript SDK and provides no headroom command — and upstream notes that clients which do not inherit an interactive shell need the absolute path to the headroom executable. Only the MCP server mode is configured: headroom proxy, headroom wrap (which also registers Serena for the wrapped project), headroom learn, and the cross-agent memory store are separate upstream features. Upstream states compression runs locally, but the [ml] extra downloads the Kompress-v2-base model from Hugging Face and proxy mode still talks to a model provider; neither is part of this recipe, and no model is pinned here. The smoke check completed initialize (server headroom, reported version 1.30.0, protocol 2025-11-25) and listed the three tools; no tool was called, so compression ratios, retrieval fidelity, and token savings are unverified.
hippo-memory: Long-term memory for coding agents over MCP: a local store that strengthens memories when they are recalled, lets stale ones decay, and answers recall, remember, outcome, conflict, and status questions through 13 tools. The recipe pins the published npm release 3.2.2 in the npx form upstream documents for clients that avoid a global install; upstream also shows hippo mcp after npm install -g hippo-memory. Only the MCP launch configuration is written here: hippo init additionally inserts a Hippo block into instruction files that already exist (it never creates one), registers a daily 6:15am run through crontab or the Windows scheduler, and learns from each registered project's git history — --no-hooks, --no-schedule, and --no-learn limit that, and Codex session capture is opt-in through hippo hook. With no store present the first tool call creates the global store under ~/.hippo; a per-project .hippo/ store appears only after hippo init, and either can be committed or ignored as the project prefers. Requires Node.js 22.16 or newer and has no other runtime dependency. Upstream's own receipts section records a benchmark round whose repeat-mistake effect did not reach its pre-registered minimum, and an earlier claim it retracted. The smoke check completed initialize (server hippo-memory, reported version 3.2.2, protocol 2024-11-05) and listed 13 tools; no tool was called, so recall quality, decay behaviour, and the published benchmark figures are unverified.
reticle: Verifies the running app rather than the source: the agent drives the project's own dev server from inside a real browser and gets pass/fail/cannot-tell verdicts with the file:line behind them, for change verification, access control, accessibility, performance, and SEO checks. The recipe pins the published npm release 3.9.0 in the MCP form upstream documents (npx @reticlehq/server@<version> mcp); the fuller install paths — the one-line shell or PowerShell installer, or reticle setup mcp after a global install — additionally append blocks to CLAUDE.md, AGENTS.md, RETICLE.md, and .claude/commands/reticle.md, add a prompt hook to each agent's project config so the agent verifies before claiming completion, and keep run state under .reticle/ (--no-mcp skips the files). Anonymous usage counts are sent by default (reticle telemetry disable, RETICLE_TELEMETRY=0, or DO_NOT_TRACK=1); prompts stay redacted locally and are uploaded only after the project is linked with reticle connect or an API key in CI. Playwright is the nearest neighbour in this collection — both drive a browser, but this one is asked for a verdict instead of raw page state, which is what its own "why not Playwright" section argues. The smoke check completed initialize (server reticle, reported version 3.9.0, protocol 2025-11-25) and listed 10 tools from a temporary directory; no app, no browser, and no tool call were involved, so no flow was actually verified.
mcp-toolbox: Google's MCP server and framework for databases: prebuilt toolsets for PostgreSQL, MySQL, SQLite, SQL Server, and BigQuery, plus SQL tools, prompts, and resources defined in a tools.yaml. It stays reference-only because every useful launch needs database connection settings supplied by the project, and the custom-tool path reads a configuration file that names those sources; npm publishes a per-platform binary behind npx -y @toolbox-sdk/server --prebuilt=postgres, so a recipe would register a server that cannot answer its first query until credentials — and, for custom tools, a tools.yaml — exist, the same reason GitHub's MCP server is reference-only. Keep that file and any credentials out of version control and out of an agent's context. Upstream also documents an OCI image, release archives, and the Google Antigravity MCP store as install paths.
open-code-review: Alibaba's agent-driven code review: deterministic analysis pipelines feed a review model, and the repository also ships two agent skills (skills/open-code-review, skills/open-code-review-delegate) plus plugins for Claude Code, Codex, Cursor, Kimi, OpenCode, and QCA. It stays reference-only because the reviewer is a globally installed npm CLI (npm install -g @alibaba-group/open-code-review, command ocr) that needs an LLM provider and an API key saved through its own interactive configuration before it can review a diff, and because the catalog's skill recipes copy reviewed files without being able to install or configure that CLI — copying the skill alone would leave a project with instructions and no engine. Delegation mode runs the review through the calling agent's own model and needs no OCR key. Upstream also documents install scripts, a release binary, and a source build. Recording it neither installs the CLI nor runs a review, spends provider tokens, or uploads source.
mcp-inspector: The reference client for MCP servers, maintained by the project that maintains the protocol: npx @modelcontextprotocol/inspector opens a web UI, --cli a scriptable client, and --tui a terminal UI over stdio, SSE, and streamable HTTP, and the package also ships an mcpdo CLI. It stays reference-only because it is a development client rather than a server a project registers, and because upstream publishes it on the npm latest tag rather than pinning a version in its own instructions. Its secret store keeps OAuth tokens and stdio env: values in the OS keychain, falling back to an unencrypted ~/.mcp-inspector/secrets.json where no keychain exists. It is the tool to reach for when a recipe in this catalog does not answer the handshake; running it changes no project and installs no server.
orm-lens: Answers Django model questions over MCP from static parsing — relations, ER diagrams, migration dependencies, cascade and blast radius, index suggestions, N+1 candidates, and schema drift — with SQLAlchemy models, Alembic revisions, and FastAPI/Flask documented as preview support. Upstream states it needs no database, no runserver, and no working virtualenv, so it still reports models from a cold clone or a broken environment; the base package accepts Python >=3.9 but the mcp extra needs >=3.10, hence the --python 3.12 launcher pin. A paid cloud service exists upstream for PR review and table statistics and is not used here. The same repository also ships a VS Code/Open VSX extension, an orm-lens CLI, a pre-commit hook, and an agent skill directory; only the MCP launch configuration is written. The smoke check completed initialize and listed 13 tools against a one-model sample module, and no tool was called, so no real project was analysed.
misata: Synthetic test fixtures rather than schema or query tooling: a declaration of the facts the data must hold — monthly revenue, channel shares, balanced ledgers, foreign keys, state machines — compiles into multi-table rows returned over MCP, with a certificate a DuckDB run can re-check. The recipe pins the published PyPI release 0.9.8.1 with its mcp extra in the form upstream documents, and pins --python 3.12 for a deterministic resolution because the package accepts Python 3.10 or newer while the extra caps mcp>=1.0.0,<2.0.0. Upstream documents 15 tools and flags seed_database as the only one that writes to a database; the smoke check completed initialize (server misata, reported version 0.9.8.1, protocol 2025-11-25) and listed all 15, and no tool was called, so no dataset was generated, no certificate was re-verified, and nothing was seeded. Upstream records its limits: exactness holds for declared facts and not for arbitrary SQL or window functions, and mimic fits to real rows rather than anonymising them. The same package also ships a CLI, a pytest plugin entry point, a Streamlit studio, and a SKILL.md; only the MCP launch configuration is written here.
seleniumbase: Browser automation over MCP from a test framework rather than a wrapper around one. The server ships inside the framework's own PyPI distribution — pip install "seleniumbase[mcp]" installs the same seleniumbase-mcp script — so the recipe pins the published release 4.57.1 with its mcp extra instead of a separate server package, and pins --python 3.13 because the extra requires Python 3.10 or newer while this host's system Python is 3.9. The distinguishing property is the transport: upstream drives Chrome through SeleniumBase's pure CDP Mode, entirely over the Chrome DevTools Protocol with no WebDriver and no chromedriver in the loop. A Chrome/Chromium installation and a usable display or headless environment are prerequisites, and no browser starts until start_browser is called. The smoke check completed initialize (server seleniumbase-mcp, protocol 2025-11-25; the server reports no version string of its own) and listed 24 tools — start_browser, close_browser, get_page_info, open_url, manage_history, find_elements, get_content, get_attributes, check_if_condition, click_element, hover_action, type_text, select_option, focus_element, wait_for_condition, assert_condition, manage_cookies, manage_storage, scroll_page, manage_window, manage_tabs, solve_captcha, save_page, and run_javascript — and no tool was called, so no page was visited and no CAPTCHA was attempted. This is a general-purpose browser driver, so it overlaps playwright and reticle: pick the one whose interface fits the job, since all three attach to the same browser family.
dbhub: A database MCP server shaped around token budget rather than tool count: two default tools (execute_sql, search_objects) plus optional custom tools, instead of the per-object toolsets other database servers generate, so a whole connection costs roughly 1.4k tokens of context. The recipe pins the published npm release 1.5.1 in the stdio form upstream documents and passes --demo, which serves the bundled sample SQLite database, so the stored command answers a handshake without credentials, an existing database, or a tools.yaml. For real work the user replaces --demo with --dsn "<url>" — PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, or sqlite: paths — and adds the guardrails upstream documents (--readonly, --max-rows, --query-timeout); that swap is exactly why mcp-toolbox stays reference-only while this one can be verified end to end. Node.js 22.5 or newer is required because the SQLite path uses the built-in node:sqlite module. Upstream also ships a TOML configuration for several databases in one process, an HTTP transport for its own web workbench, a Docker image, and a Claude Code plugin; only the stdio launch configuration is written here. The smoke check completed initialize (server DBHub MCP Server, reported version 1.5.1, protocol 2025-11-25) and listed both default tools against the bundled sample data; no query was executed, so SQL behaviour, the sample content, and the read-only flags are unverified. Any DSN, user, or token belongs in the agent's own configuration and out of version control.
awesome-copilot: A large community collection of Copilot instructions, chat modes, custom agents, skills, and prompts, contributed file by file rather than shipped as one pinnable artifact, so it is recorded without a selection being copied. Pick and review the files a project needs; nothing here activates an instruction file.
skillevaluator: Framework for scoring skill quality, detecting overlap, and measuring a skill's effect on live agent runs. It is reference-only because the PyPI name is a zero-version placeholder and the documented quickstart runs from a source checkout. Installing the catalog's skills does not evaluate them; the static path needs no key, while rubric scoring and security coverage need external services or scanners.
ai-rulez: An authoring and governance layer: one .ai-rulez/ source compiles into the native instruction files, skills, and MCP configuration of the harnesses a team uses, with ai-rulez.lock pins, Sigstore signing, SBOM output, reviewer approvals, an organisation policy layer, and a read-only MCP server for reports and skill serving. It is reference-only because the npm and PyPI packages of the same name are Go installer wrappers that download a native release binary — the wrapper's own fetch of the v5.10.0 darwin/arm64 archive timed out and retried on the smoke host before any MCP handshake — so there is no registry launch command to pin; release archives, Homebrew, and the Go module are the other distribution paths. The catalog's skill and rule-pack recipes copy reviewed files rather than generate output for dozens of targets, so nothing is installed or activated here. It overlaps skills-cli and awesome-copilot as discovery and authoring tooling.
chunkhound: Reference only because the pinned launch check did not pass: the server never answered initialize in a fresh project within a 300-second window, first failing before connecting (uvx picked Python 3.10, whose scipy build does not load on arm64 macOS) and then stalling on an outbound HTTPS request at startup even with Python 3.12 pinned, with and without --no-daemon and --no-embeddings. --read-only avoids that path but requires an index built beforehand. Indexing is an explicit step, the database is created inside the project at .chunkhound/db, and semantic search or deep research needs an embedding or LLM provider that is not configured here.
codebase-memory-mcp: The documented MCP entry points at an installed native executable, and the npm, PyPI, and Go packages of the same name only verify and cache that runtime for the host, so there is no registry launch command to pin. Its installer also writes agent configuration, rules, skills, and hooks and its indexing persists to a shared cache directory with an optional daemon, which is a broader change than writing one launch configuration.
ripwire: A ranked, deterministic call graph that answers what a task touches, what an edit breaks, and which tests reach it, plus edit-time quality deltas that report only regressions across twelve kinds, ambiguous-name and forgotten-co-change detection, and a scanner for reviewing an agent skill file before installing it. It stays reference-only because the one-line install is a curl-piped script that downloads a SHA-256-verified release binary into ~/.local/bin, stages seventeen skills for every agent it detects, and can add a commit hook — a machine-level change no recipe here models — and because no npm or PyPI package of this name exists to pin; its MCP server is the optional second interface of the same binary (ripwire --mcp), so no handshake was attempted. Bodies are redacted by default while identifiers and signatures never are, and the binary works offline. It overlaps code-review-graph, repomix, and codebase-memory-mcp as code-context tooling with a CLI-first, single-binary shape.
Two code-graph servers overlap: code-review-graph and codebase-memory-mcp both index a repository into a local graph for review and impact questions, with different runtimes and scope. Enable one per project rather than both.
pyrefly: A third Rust-based Python type checker and language server next to pyright and ty. Only the pyrefly lsp subcommand is documented as LSP entry (as registered by nvim-lspconfig); there is no MCP or agent-config recipe to pin. Upstream ships monthly minor releases that may report new errors, so a project-wide first run should be expected to fail until errors are triaged with pyrefly suppress.
tombi: The only configuration-file language server recorded here: schema-driven TOML formatting, linting, and diagnostics for pyproject.toml, uv, Ruff, and pytest settings, which is the part of a Python project an agent edits as eagerly as the source. It is reference-only like the other standalone servers because it is a raw LSP binary (tombi lsp) that an editor or agent client launches rather than an MCP service, and because the useful launch command depends on the client a project already has: one Rust binary is behind the curl installer, Homebrew, winget, uvx/uv tool install, npm, and the VS Code, Open VSX, Zed, JetBrains, Helix, Neovim, and Emacs integrations. Completion and validation come from the JSON Schema Store; nvim-lspconfig registers it for toml files and roots it at pyproject.toml.
agnix: The second configuration-facing server recorded here, after tombi, but it targets the agent's own files rather than the project's. 758 rules validate CLAUDE.md, AGENTS.md, SKILL.md, .claude/rules/*.md, hooks, plugin manifests, and MCP configuration, with several hundred of them scoped per client (Claude Code 170, Codex CLI 89, Cursor 40), and agnix --fix applies the safe subset of the findings. It is reference-only for the same reason as the other standalone servers — a raw LSP binary (agnix-lsp) that a client launches rather than an MCP service — and because the useful command depends on the client a project already has: one Rust toolchain is behind npm, Homebrew, PyPI, cargo, and the VS Code, JetBrains, Neovim, and Zed integrations. The same binary also ships a CLI and an optional agnix-mcp server; neither is configured by this catalog, and no lint run, --fix, or language-server session was performed for this entry.
skillsaw: The second entry whose subject is the agent's own files, after agnix, recorded because the delivery shape differs: a Python package on PyPI (uvx skillsaw tree, uvx skillsaw, uvx skillsaw fix, uvx skillsaw baseline, plus a .pre-commit-hooks.yaml gate) rather than a Rust LSP binary. It validates structure across Agent Skills, Agent Plugins v1, Claude Code and OpenAI Codex plugins and marketplaces, Cursor, Grok Build, Z.ai ZCode, Google Antigravity, GitHub Copilot / VS Code custom agents, OpenCode, Pi, native OpenClaw plugins, APM packages, MCP server maps, MCP Registry server.json, Vercel skills lockfiles, CLAUDE.md/AGENTS.md/GEMINI.md/QWEN.md, hooks and eval formats; applies content and token-economy rules (instruction drift across duplicate files, instruction budget, section length, weak language, contradictions), discovery-integrity checks and supply-chain rules (hooks.json remote execution, curl | sh, eval obfuscation, strict MCP allowlists, invisible unicode, prompt-injection payloads, embedded credentials); and separates deterministic autofixes from guidance. It stays reference-only as a CLI linter rather than one of the four install types — its own npx skills add stbenjam/skillsaw path installs only the skillsaw-onboard skill, which drives the same uvx CLI. It is adjacent to cross-code-organizer, skillevaluator and SkillSpector rather than a duplicate of any of them. Nothing was installed and no repository was linted.
deptry: Reference only. It is a CLI run inside a project environment, because it inspects the declared dependencies and compares them against the imports actually used; running it in the wrong interpreter reports misleading results. It is not part of the catalog's checks, and recording it neither installs nor runs it.
Conflict to resolve yourself: pyright, pyrefly, and ty are alternative Python type-checking servers, and ruff server covers lint/format diagnostics. Pick the servers your editor should run instead of enabling all of them.
Test dependencies: hypothesis is a project dependency next to pytest, not an MCP server or agent skill. Recording it neither adds nor runs tests and does not configure CI.
agent-skills collection: The two selected skills come from a larger, mostly single-file skill collection with its own plugin manifests and hooks. Only the pinned skill directories are copied; hooks, commands, and agent definitions stay upstream.
python-dev-skills: A Claude Code plugin marketplace (/plugin marketplace add wdm0006/python-skills, then /plugin install python-library-complete@dev-skills or a narrower bundle such as python-library-foundations, python-library-quality, python-web-app or python-mcp-servers) whose Python skills cover the library lifecycle — project-setup, packaging, testing-strategy, code-quality, api-design, documentation, release-management, security-audit, performance, lazy-dependencies, library-review, llm-features, mcp-servers, cli-development, web-app-architecture — with twenty shared skills/common/ skills (git hygiene, GitHub Actions, CI permissions, rendering untrusted content, destructive operations, UI design) and smaller Go, Rust, Scala, Swift and browser-extension sets. Its README states a repository rename is planned now that the collection is polyglot. It is reference-only because a bundle's members share skills/common/* files, which this catalog's recipe (one pinned subtree with a SKILL.md at its root and an archive SHA-256) cannot install without dangling references, and because the marketplace path itself is scheduled to change. Nothing was installed and no bundle was activated.
agents-cli: The first entry here about building the agent rather than tooling a project, and the ADK counterpart to the skill collections above: seven Agent Skills (google-agents-cli-workflow, google-agents-cli-adk-code, google-agents-cli-scaffold, google-agents-cli-eval, google-agents-cli-deploy, google-agents-cli-publish, google-agents-cli-observability) cover the ADK Python API, scaffolding, evaluation, deployment to Agent Runtime / Cloud Run / GKE, Gemini publishing, and Cloud Trace and Cloud Logging observability, next to the google-agents-cli PyPI package (1.13.0, Python >=3.11) whose agents-cli commands also work standalone. The skills install through the same npx skills add google/agents-cli CLI recorded as skills-cli, with plugin manifests for Claude Code and Gemini extensions. It stays reference-only because the delivered unit is a seven-skill bundle plus a generator CLI rather than one pinnable subtree or MCP launch command, and because deployment, publishing and hosted evaluation need the user's own Google Cloud project and credentials while only local create/run/eval works with an AI Studio API key; upstream also states that its usage telemetry logs a structured entry into the user's own Google Cloud project. Recorded as a discovery source: nothing was installed, no project was scaffolded and no agent was deployed.
brooks-lint: Six portable review modes — architecture audit, technical debt, health scoring, review, sweep, and test weakness — that report book-derived risk codes and a 0-100 score, packaged for fourteen coding agents. It is reference-only as a delivery boundary rather than a scope one: the documented installers (npx brooks-lint install <platform>, the shell installer, a plugin marketplace, or a manual cp -r skills/*) copy the six skill directories flat into the client's skills directory, and those skills load a shared skills/_shared/ directory. This catalog's skill recipe pins exactly one subtree and requires a SKILL.md at that subtree's root, so pinning skills/ is impossible and pinning a single mode would install a directory whose ../_shared/*.md references dangle. Upstream states its automated checks verify file layouts and resources rather than runtime behaviour on every host. It overlaps code-review-and-quality, open-code-review, and graphify in intent and is recorded for the dependency-aware review shape, not as another generic reviewer; nothing was installed and no review mode was run here.
Security tooling: Semgrep, Bandit, pip-audit, OSV-Scanner, Gitleaks, and SkillSpector remain reference-only. Installing security skills/rules does not install or run scanners, configure CI, or establish an audit result. OSV-Scanner overlaps pip-audit for Python dependencies while also covering other ecosystems and containers, so a project needs one of them, not both by default. SkillSpector is the one entry here that scans the agent tooling instead of the project — a skill directory, an MCP server, a Git URL, an archive, or an MCP registry snapshot — with static pattern checks that run offline and LLM stages that need a provider key and network access.
Rule scope: Choose the language/framework in the project. FastAPI rules are separate from general Python rules. Embedded instruction blocks have a size guard; the full set of rule packs is larger than 32 KiB.
Upstream plugin boundaries: Superpowers and ECC integrations install only the selected skills/rules. Plugin registration, agent definitions, hooks, commands, memory services, models, and account settings are outside this migration.
graphify: Registered by its own CLI rather than copied by this installer: the graph engine is the PyPI package graphifyy (uv tool install graphifyy or pipx install graphifyy, Python 3.10+), and graphify install detects the assistants present on the machine and wires the /graphify skill, with --project and --platform for a narrower registration. Copying the skill alone would leave a project with graph query instructions and no engine, the boundary already recorded for open-code-review; the registration also writes agent configuration the way ripwire does. Its distinguishing property is determinism rather than retrieval: local AST parsing over source, documentation, SQL schemas, configuration and PDFs, every edge explained with its origin, and no vector store or embedding provider. The repository carries the engine package, docs/how-it-works.md, docs/docker-mcp-sqlite.md and benchmarks; the license is Apache-2.0 with an additional MIT file, and the default branch is named v8.
codeboarding: Produces architecture documentation instead of answering tool calls: pipx install codeboarding --python python3.12, then codeboarding-setup and codeboarding full --local <repo>, which writes .codeboarding/analysis.json with nested component diagrams, and --render md|html|mdx|rst for per-component documentation; codeboarding-render re-renders a stored analysis without an LLM. It stays reference-only because the hybrid static-plus-LLM analysis needs a configured model provider (OpenAI, Anthropic, Google, Vercel, Ollama, OpenRouter or an OpenAI-compatible gateway) and sends code excerpts to it, while only an existing analysis can be rendered without credentials. Upstream also ships a VS Code/Open VSX extension, a GitHub Action that publishes architecture changes on pull requests, and a web viewer that parses local analysis files in the browser.
mcp-context-forge: A gateway rather than a server a project registers: the documented launch is uvx --from mcp-contextforge-gateway mcpgateway --host 0.0.0.0 --port 4444 after python3 -m mcpgateway.scripts.init_secrets generates JWT_SECRET_KEY and AUTH_ENCRYPTION_SECRET, which upstream requires in every environment including local development. It speaks HTTP, JSON-RPC, WebSocket, SSE and streamable HTTP, with stdio transport only for server-side use, so a written stdio configuration would register a server that cannot answer its first call - the same reason mcp-toolbox and GitHub's MCP server are reference-only. What it adds is federation and governance: a registry in front of MCP, A2A and REST/gRPC servers, virtual servers, plugins, OAuth/JWT/basic auth with conditional passthrough of upstream authorization headers, rate limiting, retries, health checks, metrics and an admin UI.
ruler: Applies one .ruler/ directory (AGENTS.md and ruler.toml, with nested scopes and a global fallback) to the native files of more than thirty coding agents through the npm package @intellectronica/ruler (ruler apply, or npx @intellectronica/ruler apply). It writes each target's own files - AGENTS.md, CLAUDE.md, .amazonq/rules/ruler_q_rules.md, firebender.json and others for instructions; .mcp.json, .cursor/mcp.json, .codex/config.toml, .gemini/settings.json, opencode.json and similar for MCP servers; and the per-agent skill and subagent directories where an agent has them - and automates .gitignore for the generated paths. It stays reference-only because one run rewrites many native files across a project rather than writing one launch configuration or copying one reviewed skill directory. It overlaps ai-rulez: that entry authors, locks, signs and policy-checks one source, while this applies an existing source to every harness in a project.
spec-kitty: Repo-native spec-driven delivery: specifications, work packages, acceptance criteria, review state, and merge decisions live in the repository, each work package runs through lanes (planned, in_progress, for_review, approved, done) in its own git worktree under .worktrees/, and the loop is specify -> plan -> tasks -> next -> review -> accept -> merge, with spec-kitty agent tasks status for progress. It is reference-only for two reasons: spec-kitty init installs the slash commands and skills into every supported harness it detects on the machine (Claude Code, Codex, Cursor, Gemini, Copilot, Windsurf, OpenCode, and Codex/Vibe/Pi/Letta/Antigravity as Agent Skills), which is a machine-level change the catalog's four install types do not model, and there is no single revision to pin — the README's Team Kitty 4.x prerelease installs spec-kitty-cli==4.0.0rc2 and states that stable launch acceptance is still pending, while the stable PyPI release is 3.2.7 (pipx install spec-kitty-cli, Python 3.11+). It overlaps openspec and spec-kit in intent and is recorded for the repo-native state, lanes, and worktree isolation rather than as another prompt template. Nothing was installed and no specification was created.
cross-code-organizer: The first entry here whose subject is the harness inventory itself rather than the project: it reports what each of Claude Code, Codex, OpenCode, and DSH will actually load per project (the effective set, not only the files on disk), accounts for context budget, edits Memory/CLAUDE.md/SKILL.md markdown in place, repairs duplicates reversibly, copies skills between harnesses, and scans MCP tool descriptions for poisoning with findings that link back to the offending server entry. It stays reference-only because it is an interactive local application — npx @mcpware/cross-code-organizer on Node.js 20 or newer, serving 127.0.0.1:3847 — rather than a skill directory, rule pack, MCP launch configuration, or LSP bridge, and because its npm latest tag still points at 0.19.3 while its own README documents 0.20.0, which is the version skew to re-check before trusting an audit result. It reads ~/.claude/, $CODEX_HOME, ~/.config/opencode/, $DSH_HOME, and discovered project directories, uploads no configuration or content, and sends only an opt-in monthly activity signal containing the UTC month, a rotating pseudonymous ID, its version, and a harness ID. It is adjacent to mcp-inspector (protocol debugging) and mcp-context-forge (gateway and registry) rather than a duplicate of either. Nothing was installed or launched.
squirrelscan is installable because its CLI publishes a pinnable npm package whose mcp subcommand runs the local stdio server (Node.js >=18): the pinned npx -y squirrelscan@0.0.173 mcp was executed through this repository's smoke script and negotiated protocol 2025-11-25, reporting squirrelscan 0.0.173 with 15 tools (audit and quick check, report and audit listing, issues, rules, entities and entity graphs, comparison, findings, feedback). It is recorded next to playwright, reticle and seleniumbase for the other half of the Web workflow: those drive or inspect a page that is still being built, this one audits a deployed site and returns findings written for an agent to fix, each pointing at the rule that produced it. The hosted endpoint (https://mcp.squirrelscan.com/mcp, HTTP transport) and the Claude Code plugin marketplace route stay upstream setup rather than part of the recipe, and the report, publish and comment scopes need the user's own account. Nothing was audited: no site was fetched, no report was written and no MCP tool was called.
pipelock is the runtime counterpart to the static scanners already here. gitleaks, SkillSpector, semgrep and osv-scanner find secrets, unsafe content and vulnerable dependencies at rest; pipelock sits inline between the agent and the network (including the MCP path), scans outbound requests and inbound responses, and mediates egress against a policy file. Its README describes an ordered URL scanner pipeline (scheme, CRLF and path-traversal checks, allow/block lists, immutable SSRF and DLP floors, path/query/subdomain entropy, nested URL destinations, DNS SSRF and rebinding protection, per-domain rate limits and data budgets) with DLP running before DNS resolution, 65 built-in DLP patterns with checksum validation, 34 prompt-injection and state/control-poisoning response patterns with six-pass normalization for zero-width characters, homoglyphs and leetspeak, and mediator-signed action receipts so a decision can be verified outside the agent runtime; pipelock demo --receipts-dir ./out produces those receipts offline with no account. It stays reference-only because it is a Go binary distributed through source builds, release archives, containers, an attested Helm chart and a GitHub Action (action.yml) rather than one of the four install types, and because running it means writing a policy and wiring a proxy for each client. Caveats: the quick start in the README targets v3.8.0 and says to use those examples after the release is published, while the latest published release checked in this pass is v3.7.1, and upstream states SVID/AARP workload-identity appraisal is verifier-side today. Nothing was installed: no proxy was started, no policy was loaded and no traffic was intercepted.
elephant-goldfish is a plugin rather than a single skill, which is why it has no recipe: ten reviewer prompts live in agents/, and the shipped skills (skills/egm-*/SKILL.md, mirrored under distributions/claude/skills/ for bootstrap, brainstorm, draft, review, review-loop, resolve and drift-report) read the repository-level prompts and standards. Upstream states the constraint itself: "Install the complete plugin so its skills can read the reviewer prompts and document standards." It ships per-harness manifests and rule files for Claude Code (.claude-plugin/), Codex (.codex-plugin/), Cursor (.cursor-plugin/ plus rules/egm.mdc), Copilot (com.github.copilot/), Antigravity/Gemini (.agents/, .gemini/) and a generic .plugin/, with per-client setup in docs/ (harnesses.md, claude.md, codex.md, cursor.md, copilot.md, opencode.md); the Claude route is claude plugin marketplace add google/elephant-goldfish-model then claude plugin install egm@egm. It is recorded in spec-driven-development next to openspec, spec-kit and spec-kitty for a different lever: instead of generating and gating specification documents, it runs an agent ensemble that drafts a design and then attacks it, and reports drift afterwards. Repository metadata, the README, the plugin manifests, the agents/ and skills trees, AI_POLICY.md and the Apache-2.0 LICENSE were read in this pass; nothing was installed and no plugin, agent or skill was activated.
lspyx covers the one gap left in lsp-bridge: serena and lsp-mcp-server bridge to configured language servers, agent-lsp orchestrates several, and mcpls bridges Rust, Python and Go - none of them is Python-only and none reports through ty. lspyx mcp serve answers definition, references, outline, diagnostics and rename-preview queries from ty plus Ruff with a per-workspace daemon (src/daemon/, src/mcp/, src/lsp/rename.rs), and the CLI exposes the same answers; setup is brew install iyazerski/tap/lspyx, cargo install lspyx --locked or install.sh, with ty and Ruff in the project .venv or on PATH, and lspyx doctor checking the result. There is no registry package the installer can own, so it stays reference-only like the other two bridges. Caveat worth weighing: 0 stars observed on 2026-11-09 - the least-adopted entry recorded here - about seven months old, single-maintainer, MIT-licensed, with eight tagged releases (latest v0.6.0, 2026-11-07); re-check it at the next review and drop it if the lane stays unserved. Nothing was installed and no MCP session was started.
planning-with-files is the first installable skill here whose state lives in files rather than in context: .planning/<date>-<slug>/ holds the plan, its progress and its findings, the active plan is re-injected every turn so it survives context rot, the skill rebuilds the plan after /clear or compaction, and scripts/check-complete.sh gates completion with a SHA-256 attestation of the plan text it injected. The recipe copies the pinned skills/planning-with-files subtree (SKILL.md, scripts/ with bash, PowerShell and Python implementations, templates/, evals/, examples.md, plan-inventory.md, reference.md), and every path in SKILL.md resolves inside that directory (${CLAUDE_SKILL_DIR}, $PWF_SKILL_DIR, ./scripts/..., no ../ escapes), so the installed skill is self-contained; the Claude Code and Codex hooks it needs are declared in the skill's own frontmatter, which those clients read from the installed file, so apd owns the files while hook registration stays upstream. Upstream additionally ships an npm launcher, plugin-marketplace entries and an npx skills route for 60+ agents; this entry deliberately pins the commit archive only. Evidence read in this pass: repository metadata, README, the skills/ tree at the pinned commit, SKILL.md, the scripts/, templates/ and evals/ listings, and the downloaded archive (24,950,981 bytes, SHA-256 verified, 37 files under the subtree, LICENSE at the archive root copied as license_path). Nothing was installed and no agent session was started.
ui-ux-pro-max covers the Web half of this catalog's remit from the design side: it answers style, palette, typography, icon, motion and UX-guideline questions from a local data set of 39 CSV/JSON files (3,103,749 bytes) under .claude/skills/ui-ux-pro-max/data/, covering 22 stacks, through scripts/search.py and design_system.py, which need only a Python 3 standard library and make no network calls, and it can emit a project design system instead of a palette. It is reference-only on purpose: upstream delivers it through a per-client generator (npm install -g ui-ux-pro-max-cli, uipro init --ai <client>, also npx, with update/uninstall/--dry-run and a Claude Code marketplace route) and its README says to always use that CLI, so the repository tree is the generator's source template rather than the installed layout; the commands inside SKILL.md are anchored to ${CLAUDE_PLUGIN_ROOT}/.claude/skills/ui-ux-pro-max/scripts/search.py, which a bare commit-archive copy would not reproduce. That, plus the CLI owning activation and updates, is why no pinned recipe is claimed and the entry sits next to skillsaw and graphify as a reference. MIT license at the repository root (skill.json also declares MIT). Evidence read in this pass: repository metadata, README (install and troubleshooting sections, including the pre-v2.5.1 symlink packaging note), skill.json 2.15.0, the .claude/skills/ trees at commit 4f2dbb67, SKILL.md and its path references. Nothing was installed and no client files were generated.
agent-skills-spec is the normative document behind the skill kind used throughout this catalog: docs/specification.mdx defines the SKILL.md frontmatter (name, description, optional license, allowed-tools and metadata), the directory layout and the progressive-disclosure rules that the skill recipes here follow, and skills-ref/ is the Python reference implementation (skills-ref validate <dir>, read-properties, to-prompt; importable API under src/, tests alongside, PyPI release 0.1.1, Python >=3.11). It is recorded as a reference next to anthropics/skills and agents-cli because it is not a tool this installer puts into a client project — it is what a skill is checked against, so there is no pinned recipe. Upstream marks skills-ref as demonstration-only rather than production tooling, which is the caveat recorded with the entry; the specification itself is the reason the entry is useful. Evidence read in this pass: repository metadata, README, the docs/ and skills-ref/ trees, skills-ref/README.md, skills-ref/pyproject.toml and the PyPI release metadata. Nothing was installed and no skill was validated with it.
arcade-mcp is the third entry in the build-mcp lane next to fastmcp and the official mcp-python-sdk, and the most widely deployed of the three (its README attributes 7,500+ tools across 81 hosted MCP servers to it): a decorator API covering tools, resources, prompts, sampling, elicitation, progress and logging, requires_auth= declarations that let the framework own OAuth and token refresh, and an arcade-mcp CLI for scaffolding, running and deploying servers. It stays reference-only for the same reason as fastmcp: it is a Python library (monorepo layout with libs/arcade-core, libs/arcade-mcp-server, libs/arcade-cli, libs/arcade-evals, libs/arcade-serve and examples/mcp_servers/) installed inside a server project rather than an agent-side MCP server the installer can configure by launching a command, so there is no pinned recipe; pip install arcade-mcp resolves PyPI 1.17.6 (Python >=3.10) and the server package publishes separately as arcade-mcp-server 1.33.0. MIT license. Evidence read in this pass: repository metadata, README (framework highlights and install sections), the libs/ and examples/ trees, and the PyPI metadata for both packages. Nothing was installed and no server was started.

MCP package/source pins identify the top-level server revision. npm/uv and upstream language-server managers may resolve additional dependencies at launch; this is not a hermetic runtime lock.

Selection, verification, and maintenance

Admission considers the coding task served, documented setup, maintenance activity, license, upstream identity, and whether a reproducible integration can be tested. The catalog separates source/configuration checks from runtime operation. Inclusion does not claim every upstream feature or every agent version was tested.

See research and verification evidence, ECC migration details, LSP and security-tool setup, and agent configuration paths.

Updates are curated and pinned. Refreshing installed instructions regenerates them from the catalog shipped with this CLI version. It does not watch upstream repositories, follow their default branches, or silently adopt new skills/rules. A catalog update requires reviewing upstream changes and licenses, selecting a new pin, recomputing archive hashes where applicable, updating these documents, and testing the installation lifecycle.

ast-grep: Structural search, linting and rewriting by AST pattern rather than by text, which is the difference that matters when an agent is asked to change every occurrence of one code shape across a project. It stays reference-only because the delivered unit is a native CLI built from a Rust workspace (crates/: cli, core, language, config, dynamic, lsp, outline, napi, pyo3, wasm) and not one of this catalog's install types: the Python route is the published package ast-grep-cli (root pyproject.toml, built with maturin from crates/pyo3), which belongs in a project environment next to ruff and deptry rather than in a client configuration, and the CLI's own lsp subcommand (crates/cli/src/lsp.rs) serves ast-grep rule diagnostics to an editor rather than the Python language intelligence the LSP entries here provide. Python patterns are parsed by crates/language/src/python.rs. Nothing was installed and no search or rewrite was run.
claude-plugins-official: The directory Anthropic maintains for Claude Code plugins: .claude-plugin/marketplace.json lists 317 plugins — skills, agents, hooks, commands and MCP servers — installed with /plugin install <name>@claude-plugins-official, with 14 partner entries under external_plugins/ and the first-party ones grouped under plugins/. It is reference-only because a marketplace is a catalogue of pointers: each plugin resolves to its own repository, subdirectory or version tag, so there is no single subtree to pin and verify — the same reason recorded for trailofbits/skills and anthropics/skills. Use it to choose a plugin, including the MCP servers already recorded here (serena, context7, playwright, github), and note the README's own boundary: Anthropic does not control the MCP servers, files or software inside third-party plugins and cannot verify they behave as intended. Nothing was installed and no plugin was activated.
mantis: Agentic vulnerability research as an orchestrated method rather than another scanner: skills/ holds 19 skills (mantis-threat-model, mantis-history, mantis-structural-index, mantis-plan, mantis-researcher, mantis-reproduce, mantis-patch, mantis-critic, mantis-dedupe, mantis-calibrate, mantis-chain, mantis-reflect, mantis-report, mantis-advise and the rest) and reference/ is a Google ADK harness that runs the same pipeline over a target checkout, with a second copy of the method inside the harness and an action/ variant for pull requests. It stays reference-only because the useful unit is that collection plus a harness needing a model provider (Vertex AI or an API key) and a target repository, not one pinnable skill directory an installer can copy — and because upstream states that findings and patches produced this way must be verified by a security expert before being reported and that the pull-request scan is untested end to end and must not gate merges. Recorded in security next to trailofbits-skills as the orchestration layer in front of the scanners already listed there. Nothing was installed, no campaign was run and no code was scanned.
retentioneering: Product analytics where this catalog previously had schema and migration tooling but no lane for user behaviour. SKILL.md plus its bundled references/ (analysis recipes, API map, gotchas and validation) and scripts/inspect_event_log.py take an agent from user-level event logs to paths, transitions, funnels, segments, retention, step matrices and sequence patterns, and every path the skill names resolves inside its own directory, so the pinned archive recipe copies a working unit. It is the only installable addition in this pass; the same repository also publishes the retentioneering Python package (5.3.0, Python >=3.11) and an MCP server that the skill calls, and the recipe installs only the skill, so the environment it analyses data in still needs the library. Nothing was installed and no dataset was analysed.
aoci-code: A local-first Go MCP server that keeps the agent's own map of a repository — one line per file in aoci.txt, with aoci.code.txt and aoci.meta.txt alongside, committed to Git and re-verified against the tree as work lands — so an agent reads the index instead of re-reading the codebase; it is the same problem codebase-memory-mcp, chunkhound and code-review-graph address by indexing for retrieval, with a file the model writes and Git versions instead. It is reference-only because delivery is a release binary (go.mod, Makefile, cmd/, engine/, internal/, .goreleaser.yml) that a project registers itself, and because the repository carries the Functional Source License 1.1 with an MIT future license (LICENSE; the GitHub API reports NOASSERTION for the adapted header), which permits non-competing use today and converts to MIT later — a second reason nothing from its spec/, docs/ or AGENTS.md text is copied into this repository. Upstream also ships NOTICE, PATENTS, TRADEMARKS, THIRD-PARTY-NOTICES, a Glama listing and bilingual READMEs. Nothing was installed and no index was built.
semble: Local code search built for agents, delivered as a stdio MCP server behind the documented semble[mcp] extra, so an agent asks a question in natural language and gets ranked snippets back instead of grepping and reading whole files. Everything runs on CPU: no GPU, no API key and no external service, which is what makes rebuilding an index cheaper than caching one inside the project — the mechanism that separates it from chunkhound and codebase-memory-mcp (embedding index kept in the project), code-review-graph (tree-sitter graph), repomix (one packed file) and aoci-code (agent-written map). One installer offers four delivery modes — MCP configuration, AGENTS.md/CLAUDE.md instructions, a semble-search sub-agent, or a semble skill plus a one-line pointer — and this recipe writes only the MCP launch. The retrieval-quality, index-speed and token-count figures on the project page (#benchmarks, assets/images/*.png) are upstream's own measurements and were not reproduced. The smoke check completed initialize (server semble, version 0.7.3, protocol 2025-11-25) and listed 2 tools (search, find_related); no query was run, no index outside the working directory was built, and neither the CLI, the sub-agent nor the skill was installed.
archy: Architecture assertions over the Python import graph — what an edit reaches, which tests it affects, where the cycles are, how coupled the modules are and whether the graph still agrees with the layers the project declared — exposed twice: a CLI whose check exits non-zero (usable in CI and pre-commit) and a stdio MCP server a session can call before it commits. Analysis is static, so it needs no interpreter run, no database and no test execution, and it writes no state into the project. It fills the lane the other Python checkers leave open: ruff and ty/pyright cover style, types and imports inside a file, while transitive reach and import direction are only visible against a declaration the project supplies. Upstream is explicit about the boundary — it is not a code-navigation tool, and that lane stays with serena, lsp-mcp-server and the descriptive graph entries — and it publishes its own negative results in docs/WHAT_DIDNT_WORK.md: across 25 unaided agent runs on the riskiest SWE-bench tasks no run introduced a cycle or a declared-layer violation (95% upper bound 12%), and 0.5% of 1,072 human commits did, which is why the README retracts the "agents will rot your import graph" premise rather than selling it. Recorded from a young repository (8 stars, last pushed 2026-11-12) whose CI, install e2e workflow, docs/SCORING.md, Claude Code marketplace manifest, bundled skills/archy skill and GitHub Action were read, not audited. The smoke check completed initialize (server archy, version 0.55.1, protocol 2025-11-25) and listed 14 tools (archy_check, archy_impact, archy_cycles, archy_graph, archy_dsm, archy_conventions, archy_brief, archy_score and the rest); no tool was called, so no project beyond the one-file sample was analysed, no layer configuration was written, and the contracts extra (import-linter) was not installed.
acumen: Skills that are built and scored, not just reviewed — the generating counterpart of the auditing tools already recorded (skillevaluator, skillsaw, skillspector): it scaffolds a configuration and a task set for one Python package, runs coding agents against the training split, rewrites the package's SKILL.md from the failures that were measured, and evaluates each new version on held-out validation tasks. It is reference-only because the delivered unit is that harness rather than one skill directory, rule pack, MCP server or bridge: using it means owning a driver model, an agent CLI, task data and throwaway environments for the target project, and the package is young (PyPI 0.0.2, Python >=3.12) with its method, configuration and validation split documented on Read the Docs. It comes from the same upstream as the Python scientific tooling this catalog already references (pytest, hypothesis). Nothing was installed, no agent was run and no skill was rewritten.

Navigation Menu

Repository navigation

CATALOG.md

CATALOG.md

Curated catalog / 精选目录

Browse and select

Presets

Engineering skills, standards, and delivery

Python, languages, and framework practices

Security review and scanning

Portable agent rules

Documentation, browser, and MCP ecosystem

Language intelligence and LSP

Specification-driven development

Skill ecosystems and discovery

Python project toolchain

Building MCP integrations

Source pins and installation boundaries

Selection, verification, and maintenance

FilesExpand file tree

CATALOG.md

Latest commit

History

CATALOG.md

File metadata and controls

Curated catalog / 精选目录

Browse and select

Presets

Engineering skills, standards, and delivery

Python, languages, and framework practices

Security review and scanning

Portable agent rules

Documentation, browser, and MCP ecosystem

Language intelligence and LSP

Specification-driven development

Skill ecosystems and discovery

Python project toolchain

Building MCP integrations

Source pins and installation boundaries

Selection, verification, and maintenance