Semantic Cache MCP

Cut your MCP client's token usage by ~98% on cached reads, with millisecond responses.

Semantic Cache MCP is a Model Context Protocol server that puts every file operation behind one cache. Re-reading a file you already hold costs a few tokens instead of the whole file, and search and grep run over that same corpus rather than the disk.

Fourteen tools share the layer: read, read_image, batch_read, warm, write, edit, edit_preview, batch_edit, search, grep, glob, delete, clear, stats.

Why this exists

Reads stop costing tokens. The first read hands back a content_hash. Send it back — known_hash on read, a known_hashes entry on batch_read — and the server replies unchanged without resending. A modified file returns a diff with changed line numbers; an oversized one collapses to a structure-preserving summary rather than a blind cut at a byte offset.

Hashes travel as their first 16 hex characters — a claim is only ever checked against the entry for the path it names, so 64 bits separates two versions of one file with room to spare, and the full digest is still accepted. A shorter prefix is not: that would match every version at once.

That echoed hash is the whole contract, and it is the only evidence the server has that a file is still in your context. A warm cache proves the server holds the file, never that you do — the store is on disk and outlives the process, the session, and your context window. A read without a matching hash always sends the file, so forgetting is safe: after a compaction, omit the hashes and get your files back in full.

Search and grep run on the cache, not the disk. BM25 keyword search, glob, and grep all read the corpus that read, batch_read and warm populate — and warm fills it without returning a byte of content, so a whole tree becomes searchable for a few dozen tokens. An in-session result LRU collapses repeated queries to sub-millisecond hits.

Mutations are bounded by default. write, edit, and batch_edit enforce size and match limits, can run formatters, and refresh the cache atomically. A dry_run writes nothing and says so — the status becomes would_create / would_update / would_edit — so a preview is never mistaken for a completed write.

Installation

Add to Claude Code settings (~/.claude.json).

Option 1: uvx, always runs the latest version:

{
  "mcpServers": {
    "semantic-cache": {
      "command": "uvx",
      "args": ["semantic-cache-mcp"]
    }
  }
}

Option 2: uv tool install:

uv tool install semantic-cache-mcp

{
  "mcpServers": {
    "semantic-cache": {
      "command": "semantic-cache-mcp"
    }
  }
}

Restart Claude Code.

Block Native File Tools (Recommended)

Disable the client's built-in file tools so all file I/O routes through semantic-cache.

Claude Code — ~/.claude/settings.json:

{
  "permissions": {
    "deny": ["Read", "Edit", "Write"]
  }
}

OpenCode — ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "read": "deny",
    "edit": "deny",
    "write": "deny"
  }
}

CLAUDE.md Configuration

Add to ~/.claude/CLAUDE.md to enforce semantic-cache globally:

## Tools

- MUST use `semantic-cache-mcp` instead of native I/O tools (98% token savings on cached reads)

Tools

Core

Tool	Description
`read`	Cache-aware single-file read: full content plus a `content_hash` on the first read, `unchanged` for a matching `known_hash`, a diff for a changed file. `offset`/`limit` recover exact line ranges, with the number gutter opt-in via `line_numbers=true` (it costs ~17% of a window, and the range is in `lines` regardless). `outline=true` returns one `line: signature` per definition instead of the text — the cheap first read of a large file, and a map rather than possession, so it reports `file_hash`. A partial or summarized read reports `file_hash` (prefixed `partial:`) — it identifies the file but is never proof you hold it. A ranged read also returns a signed `coverage_token` for the lines delivered: echo it back and a window you hold answers `unchanged`; windows covering the whole file mint a claimable `content_hash`.
`read_image`	Image pass-through. Returns an MCP image content block (base64 + mime) so vision models see the pixels; sidecar metadata carries size and mime. Format verified by magic bytes (PNG, JPEG, GIF, TIFF, BMP, WebP), not extension. Bypasses the cache. Capped at 5 MiB (`SCMCP_MAX_IMAGE_BYTES`).
`write`	Full-file create or replace with cache refresh. Returns creation status or an overwrite diff; supports `append=true` and formatters. A full write hands back a claimable `content_hash`; an append needs `known_hash` to earn one.
`edit`	Exact edit against cached content, with scoped and line-range modes plus `dry_run=true`. Pass `known_hash` to get a claimable `content_hash` back and skip the read afterwards. For several edits to one file, use `batch_edit`.
`batch_edit`	Many exact edits to one file, applied atomically, with per-edit success reporting. Takes `known_hash` on the same terms as `edit`. An ambiguous anchor, an anchor inside another edit's line range, and two overlapping ranges are each rejected rather than silently resolved; every reported success is verified against the text it produced.
`edit_preview`	Read-only probe returning match count, line numbers, and context snippets for a candidate `old_string`. Confirms anchor uniqueness before a costly `edit`.
`delete`	Single-path delete for a file or symlink, with cache eviction and `dry_run=true`. No globs, no recursion, no directory delete.

Discovery

Tool	Description
`warm`	Index files into the cache so `grep` and `search` can see them, returning counts only — never content. Takes paths or globs. Every file left out is reported with a reason (`not_found`, `not_a_file`, `binary`, `too_large`, `unreadable`, `timeout`), and a cap that stops the walk sets `truncated` or `incomplete` rather than a short count that reads as complete.
`batch_read`	Multi-file cache-aware read. Handles globs, priorities, token budgets, and diff/full routing. Returns each file's `content_hash`; pass them back as `known_hashes` to suppress the ones you still hold.
`search`	Cache-only BM25 ranking of cached files. Terms join with `OR`, so a word your corpus lacks narrows the ranking instead of emptying the results. Previews centre on the matching term rather than the file's first 200 characters, so they show why the file ranked. Index likely files with `warm` first.
`grep`	Cache-only exact search — regex or literal. Best for symbols and exact strings. Hits come back as `"<line>:<text>"` strings grouped by file, context lines as `"<line>-<text>"`, with overlapping context windows merged so no line is sent twice; paths are relative to the `root` the response names. `output="paths"` answers "which files mention X" and `output="count"` just the totals. An invalid, over-long, or catastrophically backtracking pattern is an error, never an empty result; use `fixed_string=true` for literal text. Responses state whether the scan completed, so a capped result is never read as a total.
`glob`	File discovery plus cache coverage. Find candidates, then pass the paths to `warm` (to index them) or `batch_read` (to read them).

Management

Tool	Description
`stats`	Cache metrics, session usage (tokens saved, tool calls), and lifetime aggregates.
`clear`	Reset all cache entries.

Tool Reference

The table above is the authoritative map; these are the common call shapes.

read: single file, automatic caching

read path="/src/app.py"                        # automatic: full, unchanged, or diff
read path="/src/app.py" offset=120 limit=80    # lines 120 to 199 only

State	Response	Token cost
First read	Full content plus a `content_hash`	Normal
Unchanged	`unchanged: true`, when you pass back a matching `known_hash`	A few tokens
Modified	Unified diff only	5 to 20% of original

write: create or overwrite files

write path="/src/new.py" content="..."
write path="/src/new.py" content="..." auto_format=true
write path="/src/large.py" content="...chunk1..." append=false   # first chunk
write path="/src/large.py" content="...chunk2..." append=true    # subsequent chunks

edit: find/replace with three modes

# Mode A: find/replace, searches the entire file
edit path="/src/app.py" old_string="def foo():" new_string="def foo(x: int):"
edit path="/src/app.py" old_string="..." new_string="..." replace_all=true auto_format=true

# Mode B: scoped find/replace, searches only within the line range (a shorter old_string works)
edit path="/src/app.py" old_string="pass" new_string="return x" start_line=42 end_line=42

# Mode C: line replace, swaps the whole range with no old_string needed (most token savings)
edit path="/src/app.py" new_string="    return result\n" start_line=80 end_line=83

Mode	Parameters	Best for
Find/replace	`old_string` + `new_string`	Unique strings, no line numbers known
Scoped	`old_string` + `new_string` + `start_line`/`end_line`	Shorter context when `read` gave you line numbers
Line replace	`new_string` + `start_line`/`end_line`	Maximum token savings when line numbers are known

batch_edit: multiple edits in one call

# Mode A: find/replace, [old, new]
batch_edit path="/src/app.py" edits='[["old1","new1"],["old2","new2"]]'

# Mode B: scoped, [old, new, start_line, end_line]
batch_edit path="/src/app.py" edits='[["pass","return x",42,42]]'

# Mode C: line replace, [null, new, start_line, end_line]
batch_edit path="/src/app.py" edits='[[null,"    return result\n",80,83]]'

# Mixed modes in one call (object syntax also supported)
batch_edit path="/src/app.py" edits='[
  ["old1", "new1"],
  {"old": "pass", "new": "return x", "start_line": 42, "end_line": 42},
  {"old": null, "new": "    return result\n", "start_line": 80, "end_line": 83}
]' auto_format=true

batch_read: multiple files with a token budget

batch_read paths="/src/a.py,/src/b.py" max_total_tokens=50000
batch_read paths='["/src/a.py","/src/b.py"]' priority="/src/main.py"
batch_read paths="/src/*.py" max_total_tokens=30000
batch_read paths="/src/a.py,/src/b.py" known_hashes='{"/src/a.py":"8f3c..."}'

Expands simple globs, honors priority, enforces max_total_tokens, and reports skipped paths with recovery hints. Every file is returned in full unless you prove you still hold it: echo the delivered content_hash values back as known_hashes and the ones you hold collapse into an unchanged count.

warm: make a tree searchable without reading it

warm paths="src/**/*.py"
warm paths="src/a.py,src/b.py"
warm paths="src/**/*" max_files=500

Indexes the files into the cache and returns counts — warmed, already_current, skipped, tokens_indexed — with no content, no previews, and no per-file paths for the ones that worked. Anything skipped comes back under failures with a reason, and a cap that stops the walk sets truncated or incomplete.

The usual opening move on an unfamiliar tree: warm, then grep for the exact string or search for the concept, then read only what those name.

discovery: search, glob, grep

search query="authentication middleware logic" k=5
glob pattern="**/*.py" directory="./src" cached_only=true
grep pattern="class Cache" path="src/**/*.py"
grep pattern="content_hash" output="paths"
grep pattern="TODO" output="count"

A grep response names the shared directory once as root and reports each file's hits as "<line>:<text>" strings — measured at 37% fewer tokens than the per-match objects it replaced, and glob at 50%. output="count" turns a 2.6k-token answer into 77.

Configuration

Environment Variables

Variable	Default	Description
`LOG_LEVEL`	`INFO`	Logging verbosity (`DEBUG`, `INFO`, `WARNING`, `ERROR`)
`TOOL_OUTPUT_MODE`	`compact`	Response detail (`compact`, `normal`, `debug`)
`TOOL_MAX_RESPONSE_TOKENS`	`0`	Global response token cap (`0` = disabled)
`TOOL_TIMEOUT`	`30`	Seconds before a tool call times out (auto-resets executor)
`MAX_CONTENT_SIZE`	`100000`	Max bytes returned by read operations
`MAX_CACHE_ENTRIES`	`10000`	Max cache entries before W-TinyLFU eviction
`SEMANTIC_CACHE_DIR`	(platform)	Override cache/database directory path
`SCMCP_STRUCTURED_CONTENT`	`false`	Also send each result as MCP `structuredContent`. Off by default: it duplicates the text block byte for byte, and clients disagree about which they forward, so leaving it on can double the cost of every file delivered.
`SCMCP_PUBLISH_OUTPUT_SCHEMA`	`false`	Advertise per-tool output schemas in `tools/list`. Off by default: they were 11.5k of this server's 19.8k advertised tokens, paid on every request, and the Anthropic Messages API has no field to receive them. Turning this on forces `SCMCP_STRUCTURED_CONTENT` on too, since MCP requires structured content from any tool that declares a schema.

A malformed value falls back to the default and logs a warning naming the variable. See docs/env_variables.md for detail.

Safety Limits

Limit	Value	Protects against
`MAX_WRITE_SIZE`	10 MB	Memory exhaustion via large writes
`MAX_EDIT_SIZE`	10 MB	Memory exhaustion via large file edits, in `edit` and `batch_edit` alike
`MAX_MATCHES`	10,000	CPU exhaustion via unbounded `replace_all`
`GREP_MAX_PATTERN_LEN`	1,000 chars	Oversized `grep` regex source
Regex shape check	—	Catastrophic backtracking (details)

MCP Server Config

{
  "mcpServers": {
    "semantic-cache": {
      "command": "uvx",
      "args": ["semantic-cache-mcp"],
      "env": {
        "LOG_LEVEL": "INFO",
        "TOOL_OUTPUT_MODE": "compact",
        "MAX_CONTENT_SIZE": "100000"
      }
    }
  }
}

Cache location: ~/.cache/semantic-cache-mcp/ (Linux), ~/Library/Caches/semantic-cache-mcp/ (macOS), %LOCALAPPDATA%\semantic-cache-mcp\ (Windows). Override with SEMANTIC_CACHE_DIR.

How It Works

┌──────────┐     ┌────────────┐     ┌──────────────────────────┐
│  Claude  │────▶│ smart_read │────▶│ stat() + cache lookup    │
│   Code   │     │            │     │ (BEFORE any disk read)   │
└──────────┘     └────────────┘     └──────────────────────────┘
                        │
       ┌────────────────┼─────────────────┬──────────────────┐
       ▼                ▼                 ▼                  ▼
 ┌──────────┐    ┌──────────┐      ┌──────────┐      ┌────────────┐
 │ mtime    │    │ mtime    │      │ Changed  │      │ New /      │
 │ match    │    │ drift,   │      │ content  │      │ Large      │
 │ FAST     │    │ hash     │      │ → diff   │      │ → summary  │
 │ PATH     │    │ match    │      │ (80-95%) │      │  or full   │
 │ ~5 tok   │    │ ~5 tok   │      └──────────┘      └────────────┘
 │ (99%)    │    │ (99%)    │
 │ ~1 ms    │    │ ~1 ms    │
 │ no I/O   │    │ +update  │
 └──────────┘    └──────────┘

search is cached on the same principle. An in-session LRU keyed on (query, k, directory) returns warm hits in ~10 µs, and misses fall through to BM25. Every cache mutation (put, clear, delete_path, update_mtime) bumps the LRU, so callers never see a result that predates a write.

Performance

Measured on this project's 41 source files (212,499 tokens), i9-13900K, ext4 on NVMe, corpus held fixed across phases. Every phase models a caller that keeps its hashes and echoes them back — that is what earns the savings.

Token savings: 98.9% overall (phases 2 to 6)

Phase	Scenario	Savings
Overall (cached, phases 2 to 6)	Aggregate token reduction	98.9%
Unchanged re-read	mtime match, fast path skips disk I/O	99.3%
Content hash	mtime drifted, BLAKE3 still matches	99.3%
Batch read	All files via `batch_read`, 200K budget	99.3%
Search previews	5 queries × k=5, previews vs full reads	98.6%
Small edits	Real ~5% line changes in 30% of files	98.1%
Cold read	First read, no cache; one file exceeds `MAX_CONTENT_SIZE` and returns summarised, which is not a cache saving	5.9%

Latency: unchanged reads ~1 ms; repeat searches < 0.01 ms

Operation	p50	Notes
Single unchanged read (fast path)	1.1 ms	mtime + cache hit, no disk I/O
Single diff read (changed file)	0.7 ms	hash check + unified diff
Search k=5 (cache hit)	< 0.01 ms	in-session LRU
Search k=5 (cache miss)	1.4 ms	BM25 keyword search
Edit (scoped find/replace)	3.1 ms	cached content, plus the atomic write's fsync
Grep (literal `def` )	1.5 ms	FTS5 over cached corpus
Grep (regex)	3.4 ms	compiled once
Batch read (41 files, diff mode)	45.6 ms	chunk + tokenize changed files; one summarises each full pass
Unchanged re-read (41 files)	19.5 ms	whole-corpus pass
Cold read (41 files, total)	100 ms	single unrepeated pass: I/O, tokenisation, one summarisation
Write (200-line file)	2.7 ms	creates + caches, durable before it returns

Run them yourself. Pin TMPDIR to a real disk — the default /tmp is usually tmpfs, which discards fsync and reports write latency ~40% low:

TMPDIR="$HOME/.cache/scmcp-bench" \
  uv run python benchmarks/benchmark_performance.py    # operation latency
uv run python benchmarks/benchmark_token_savings.py    # token savings

See docs/performance.md for full methodology.

Documentation

Guide	Description
Architecture	Component design, algorithms, data flow
Performance	Benchmarks, methodology, cache footprint
Security	Threat model, input validation, size limits
Advanced Usage	Programmatic API, custom storage backends
Troubleshooting	Common issues, debug logging
Environment Variables	All env vars with defaults and examples

Contributing

git clone https://github.com/CoderDayton/semantic-cache-mcp.git
cd semantic-cache-mcp
uv sync
uv run pytest

See CONTRIBUTING.md for commit conventions, pre-commit hooks, and code standards.

License

MIT License. Use it freely in personal and commercial projects.

Credits

Built with FastMCP 4.0+ and:

SQLite with FTS5 for keyword (BM25) full-text search, vendored as a small built-in store
Semantic summarization based on TCRA-LLM (arXiv:2310.15556)
BLAKE3 cryptographic hashing for content freshness
W-TinyLFU frequency-aware cache eviction

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
benchmarks		benchmarks
docs		docs
scripts		scripts
src/semantic_cache_mcp		src/semantic_cache_mcp
tests		tests
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
glama.json		glama.json
lefthook.yml		lefthook.yml
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Navigation Menu

Repository files navigation

Semantic Cache MCP

Why this exists

Installation

Block Native File Tools (Recommended)

CLAUDE.md Configuration

Tools

Core

Discovery

Management

Tool Reference

Configuration

Environment Variables

Safety Limits

MCP Server Config

How It Works

Performance

Token savings: 98.9% overall (phases 2 to 6)

Latency: unchanged reads ~1 ms; repeat searches < 0.01 ms

Documentation

Contributing

License

Credits

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Semantic Cache MCP

Why this exists

Installation

Block Native File Tools (Recommended)

CLAUDE.md Configuration

Tools

Core

Discovery

Management

Tool Reference

Configuration

Environment Variables

Safety Limits

MCP Server Config

How It Works

Performance

Token savings: 98.9% overall (phases 2 to 6)

Latency: unchanged reads ~1 ms; repeat searches < 0.01 ms

Documentation

Contributing

License

Credits

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages