semantic-cache-mcp 0.3.1

Lightweight MCP server for semantic file caching with 80%+ token reduction

Semantic Cache MCP

---

Reduce Claude Code token usage by 80%+ with intelligent file caching.

Semantic Cache MCP is a Model Context Protocol server that eliminates redundant token consumption when Claude reads files. Instead of sending full file contents on every request, it returns diffs for changed files, suppresses unchanged files entirely, and intelligently summarizes large files — all transparently through 12 purpose-built MCP tools.

---

Features

• 80%+ Token Reduction — Unchanged files cost ~0 tokens; changed files return diffs only
• Three-State Read Model — First read (full + cache), unchanged (message only, 99% savings), modified (diff, 80–95% savings)
• Semantic Search — Hybrid BM25 + HNSW vector search via local ONNX embeddings (configurable model, default BAAI/bge-small-en-v1.5), no API keys, works offline
• Batch Embedding — batch_smart_read pre-scans all new/changed files and embeds them in a single model call (N calls → 1)
• Content Hash Freshness — BLAKE3 hash detects when mtime changes but content is identical (touch, git checkout) — returns cached instead of re-reading
• Grep — Regex/literal pattern search across cached files with line numbers and context
• Semantic Summarization — 50–80% token savings on large files, structure preserved
• DoS Protection — Write size, edit size, and match count limits enforced at every boundary

---

Installation

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

Option 1 — uvx (always runs latest version):

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

Option 2 — uv tool install (recommended for multiple clients):

uv tool install semantic-cache-mcp

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

Restart Claude Code. Done.

Why Option 2? — uvx spawns an isolated process per invocation, each loading its own embedding model (~200MB). If you run multiple Claude Code instances concurrently (e.g. across different projects), each one loads a separate copy, multiplying RAM usage. uv tool install puts the binary on your PATH so all projects share one installed copy and the model is loaded once per process.

GPU Acceleration (Optional)

For NVIDIA GPU acceleration, install with the gpu extra:

uv tool install "semantic-cache-mcp[gpu]"
# or with uvx: uvx "semantic-cache-mcp[gpu]"

Then set EMBEDDING_DEVICE=gpu in your MCP config env block. Falls back to CPU automatically if CUDA is unavailable.

Block Native File Tools (Recommended)

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

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

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

OpenCode — add to ~/.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` instead of native Read/Write/Edit (80%+ token savings)
  - `read` / `batch_read` → file reading with diff-mode (set diff_mode=false after context compression)
  - `write` → new files or full rewrites; `append=true` for large files
  - `edit` / `batch_edit` → find/replace (full-file / scoped / line-replace)
  - `search` / `similar` → semantic search (seed cache first with read/batch_read)
  - `grep` → regex/literal pattern search across cached files
  - `glob` → find files by pattern; `cached_only=true` to filter to cached files
  - `diff` → compare two files with semantic similarity score
  - `stats` / `clear` → cache metrics and reset

---

Tools

Core

Tool	Description
read	Smart file reading with diff-mode. Three states: first read (full + cache), unchanged (99% savings), modified (diff, 80–95% savings). Use offset/limit for line ranges.
write	Write files with cache integration. auto_format=true runs formatter. append=true enables chunked writes for large files. Returns diff on overwrite.
edit	Find/replace using cached reads — three modes: full-file, scoped to a line range, or direct line replacement. dry_run=true previews. replace_all=true handles multiple matches. Returns unified diff.
batch_edit	Up to 50 edits per call with partial success. Each entry can be find/replace, scoped, or line-range replacement. auto_format=true and dry_run=true supported.

Discovery

Tool	Description
search	Semantic/embedding search across cached files by meaning — not keywords. Seed cache first with read or batch_read.
similar	Finds semantically similar cached files to a given path. Start with k=3–5. Only searches cached files.
glob	Pattern matching with cache status per file. cached_only=true filters to already-cached files. Max 1000 matches, 5s timeout.
batch_read	Read 2+ files in one call. Supports glob expansion in paths, priority ordering, token budget, and per-file diff suppression for unchanged files. Pre-scans and batch-embeds all new/changed files in a single model call. Set diff_mode=false after context compression.
grep	Regex or literal pattern search across cached files with line numbers and optional context lines. Like ripgrep for the cache.
diff	Compare two files. Returns unified diff plus semantic similarity score. Large diffs are auto-summarized to stay within token budget.

Management

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

---

Tool Reference

read — Single file with diff-mode

read path="/src/app.py"
read path="/src/app.py" diff_mode=true         # default
read path="/src/app.py" diff_mode=false        # full content (use after context compression)
read path="/src/app.py" offset=120 limit=80    # lines 120–199 only

Three states:

State	Response	Token cost
First read	Full content + cached	Normal
Unchanged	"File unchanged (1,234 tokens cached)"	~5 tokens
Modified	Unified diff only	5–20% of original

Set diff_mode=false after context compression — Claude has lost its cached copy and needs full content.

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

• Returns diff on overwrite, confirms creation on new files
• append=true appends content rather than replacing — use for writing large files in chunks
• Cache is updated immediately after write

edit — Find/replace with three modes

# Mode A — find/replace: searches 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: search only within line range (shorter old_string suffices)
edit path="/src/app.py" old_string="pass" new_string="return x" start_line=42 end_line=42

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

Mode selection:

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 (no old_string)	Maximum token savings when line numbers are known

• Uses cached content — no token cost for the read
• Returns unified diff of the change
• Multiple matches in scope: fails with hint to add context or use replace_all=true
• Use batch_edit when applying 2+ independent changes to the same file

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

• Up to 50 edits per call — each entry can use any mode independently
• Partial success: individual edit failures don't block others
• Single round-trip, single cache update
• Failures reported per-entry so you can retry only what failed

search — Semantic search across cached files

search query="authentication middleware logic" k=5
search query="database connection pooling" k=3

• Embedding-based semantic search — finds meaning, not keywords
• Only searches files that have been previously cached via read or batch_read
• Seed the cache first, then search

similar — Find semantically related files

similar path="/src/auth.py" k=3
similar path="/tests/test_auth.py" k=5

• Finds cached files most similar to the given file
• Useful for discovering related tests, implementations, or documentation
• Only considers cached files; start with k=3–5

glob — Pattern matching with cache awareness

glob pattern="**/*.py" directory="./src"
glob pattern="**/*.py" directory="./src" cached_only=true

• Shows cache status (cached/uncached) for each matched file
• cached_only=true returns only files already in cache — useful for scoping searches
• Max 1000 matches, 5-second timeout

batch_read — Multiple files with token budget

batch_read paths="/src/a.py,/src/b.py" max_total_tokens=50000
batch_read paths='["/src/a.py","/src/b.py"]' diff_mode=true priority="/src/main.py"
batch_read paths="/src/*.py" max_total_tokens=30000 diff_mode=false

• Glob expansion: src/*.py expanded inline (max 50 files per glob)
• Priority ordering: priority paths read first, remainder sorted smallest-first
• Token budget: stops reading new files once max_total_tokens reached; skipped files include est_tokens hint
• Unchanged suppression: unchanged files appear in summary.unchanged with no content (zero tokens)
• Batch embedding: pre-scans all new/changed files and embeds them in a single model call before reading — N model calls reduced to 1
• Context compression recovery: set diff_mode=false when Claude needs full content after losing context

diff — Compare two files

diff path1="/src/v1.py" path2="/src/v2.py"

• Returns unified diff between two files
• Includes semantic similarity score (cosine distance of embeddings)
• Large diffs auto-summarized to stay within token budget

---

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)
MAX_CONTENT_SIZE	100000	Max bytes returned by read operations
MAX_CACHE_ENTRIES	10000	Max cache entries before LRU-K eviction
EMBEDDING_DEVICE	cpu	Embedding hardware: cpu, cuda (GPU), auto (detect)
EMBEDDING_MODEL	BAAI/bge-small-en-v1.5	FastEmbed model for search/similarity (options)
SEMANTIC_CACHE_DIR	(platform)	Override cache/database directory path

See docs/env_variables.md for detailed descriptions, model selection guidance, and examples.

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
MAX_MATCHES	10,000	CPU exhaustion via unbounded replace_all

MCP Server Config

{
  "mcpServers": {
    "semantic-cache": {
      "command": "uvx",
      "args": ["semantic-cache-mcp"],
      "env": {
        "LOG_LEVEL": "INFO",
        "TOOL_OUTPUT_MODE": "compact",
        "MAX_CONTENT_SIZE": "100000",
        "EMBEDDING_DEVICE": "cpu",
        "EMBEDDING_MODEL": "BAAI/bge-small-en-v1.5"
      }
    }
  }
}

Embeddings: Uses FastEmbed with BAAI/bge-small-en-v1.5 by default (33M params, 384-dimensional, 512 token context). Runs entirely locally via ONNX Runtime — no API keys, no network calls during search. Set EMBEDDING_MODEL to use a different model, and EMBEDDING_DEVICE to control hardware: cpu (default), cuda (GPU), or auto (detect available).

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

---

How It Works

┌─────────────┐     ┌──────────────┐     ┌──────────────────┐
│  Claude     │────▶│  smart_read  │────▶│  Cache Lookup    │
│  Code       │     │              │     │  (VectorStorage) │
└─────────────┘     └──────────────┘     └──────────────────┘
                           │
         ┌─────────────────┼─────────────────┐
         ▼                 ▼                 ▼
   ┌──────────┐     ┌──────────┐     ┌──────────────┐
   │Unchanged │     │ Changed  │     │  New / Large │
   │  ~0 tok  │     │  diff    │     │ summarize or │
   │  (99%)   │     │ (80-95%) │     │ full content │
   └──────────┘     └──────────┘     └──────────────┘

Read pipeline (in priority order):

1. File unchanged — mtime matches cache entry → return "no changes" message (~5 tokens)
2. File changed — compute unified diff → return diff only (80–95% savings)
3. Semantically similar cached file — return diff from nearest neighbor (HNSW vector search)
4. Large file — semantic summarization preserving docstrings and key function signatures
5. New file — full content returned and embedded; batch_read pre-scans and embeds all new files in a single model call

---

Performance

Measured on this project's 30 source files (~136K tokens). Benchmarks run on a standard dev machine (CPU embeddings).

Token Savings

Phase	Scenario	Savings
Cold read	First read, no cache	0% (baseline)
Unchanged re-read	Same files, no modifications	99.1%
Content hash	Touch files (mtime changed, content identical)	99.1%
Small edits	~5% of lines changed in 30% of files	98.1%
Batch read	All files via batch_read	99.1%
Search	5 queries × k=5, previews vs full reads	98.4%
Overall (cached)	Phases 2–6 combined	98.8%

Operation Latency

Operation	Time
Unchanged read (single file)	2 ms
Unchanged re-read (29 files)	25 ms
Batch read (29 files, diff mode)	35 ms
Cold read (29 files, incl. embed)	2,554 ms
Write (200-line file)	47 ms
Edit (scoped find/replace)	48 ms
Semantic search (k=5)	4 ms
Semantic search (k=10)	5 ms
Find similar (k=3)	49 ms
Grep (literal)	1 ms
Grep (regex)	2 ms
Embedding model warmup	206 ms
Single embedding (largest file)	47 ms
Batch embedding (10 files)	469 ms

Run benchmarks yourself:

uv run python benchmarks/benchmark_token_savings.py    # token savings
uv run python benchmarks/benchmark_performance.py      # operation latency

See docs/performance.md for full benchmarks and methodology.

---

Documentation

Guide	Description
Architecture	Component design, algorithms, data flow
Performance	Optimization techniques, benchmarks
Security	Threat model, input validation, size limits
Advanced Usage	Programmatic API, custom storage backends
Troubleshooting	Common issues, debug logging
Environment Variables	All configurable 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

This project uses Python 3.12+, strict type hints throughout, Ruff for formatting and linting, and pytest for testing. See CONTRIBUTING.md for commit conventions, pre-commit hooks, and code standards.

---

License

MIT License — use freely in personal and commercial projects.

---

Credits

Built with FastMCP 3.0 and:

• FastEmbed — local ONNX embeddings (configurable, default BAAI/bge-small-en-v1.5)
• SimpleVecDB — HNSW vector storage with FTS5 keyword search
• Semantic summarization based on TCRA-LLM (arXiv:2310.15556)
• BLAKE3 cryptographic hashing for content freshness
• LRU-K frequency-aware cache eviction