Skip to content
Octo-LexPublic

About

Local-first, privacy-first governed AI agent — built in Rust

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

455 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

OpenWand

Conjure results from intent.

OpenWand is a local-first, privacy-first AI agent desktop tool built entirely in Rust. Not a framework. Not an OS. Your personal agent — fast, small, extensible, governed, and yours.

Architecture

14 crates, one binary, zero compromise.

openwand (binary)
├── openwand-core       Agent loop, planning, recovery, mode
├── openwand-trace      Authority, append, query, stream, store
├── openwand-store      Persistence layer
├── openwand-session    Session runner, config, testing harness
├── openwand-memory     3-tier memory (user/session/agent + ACE Skillbook)
├── openwand-tools      Built-in tools (filesystem, shell, web, browser, spawn)
├── openwand-mcp-pool   MCP server pool via rmcp (official Rust SDK)
├── openwand-policy     Governance: budget governor, redaction, access control
├── openwand-llm        Multi-provider LLM routing with model cascade
├── openwand-skills     YAML + Markdown skill store with auto-discovery
├── openwand-goals      Fitness functions + autonomous improvement loops
├── openwand-content    Rich content rendering (stub — future crate)
├── openwand-workflow   Evidence ladder: 44 modules, 6 dependencies, leaf crate
└── openwand-app        CLI binary, lib, UI, persistence

Quick Start

# Build
cargo build --workspace

# Run
cargo run -p openwand-app

# Test everything
cargo test --workspace --features "openwand-session/testing,openwand-session/sqlite-testing,openwand-memory/testing,openwand-memory/sqlite-testing"

Stack

Layer Technology
UI Dioxus 0.7 (desktop)
Sessions Loro CRDT
MCP rmcp (official Rust SDK)
Memory rusqlite
Trust blake3
Runtime tokio

Principles

  1. Local-first — all data in ~/.openwand/, zero cloud dependencies
  2. Fast — compact binary, fast startup
  3. Safe — every tool call passes through policy engine before execution
  4. Observable — every action produces a structured, typed result
  5. Self-improving — agents learn from mistakes via ACE Skillbook
  6. Extensible — MCP servers add capabilities without touching core
  7. Governed — workflow evidence ladder with hash-bound append-only records
  8. Disk-verified — every development wave starts from repository truth

Workflow Evidence Ladder

OpenWand's workflow system builds an append-only evidence chain (structural hash-chaining; immutability enforcement deferred to verifier, not yet implemented):

skills/goals → task plan → review → proposal → review → readiness
→ execution → routing → bridge → outcome → reconciliation
→ run revision → continuation → next-action review → routing readiness
→ routing gate → loop controller → command composer → command review
→ manual result capture

Each step is content-addressed (blake3), hash-bound to its predecessor, and guarded against unauthorized mutation. No runtime verifier enforces append-only at the store level yet. The workflow crate is a leaf dependency with exactly 6 crates: serde, serde_json, blake3, chrono, thiserror, tracing.

Development Doctrine

From Wave 38 onward, OpenWand follows Disk-Verified Large-Wave Execution:

Scale by wave.
Ground by disk.
Accept by tests.
Seal by evidence.

No wave plan is valid until the relevant repository state has been inspected. No wave is accepted without passing tests and evidence of no forbidden mutations.

See GOVERNANCE.md, ROADMAP.md, and WAVES.md.

Test Baseline

3,999 tests (Windows), zero failures (v0.4.0 stable)

License

MIT

About

Local-first, privacy-first governed AI agent — built in Rust

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages