codeloom MCPサーバー

---

AI coding agents are powerful but fundamentally blind to your codebase structure. When your agent edits validate_token(), it has no idea that 47 callers depend on its return type. When it searches for "database connection", it greps blindly through every file. Without a code graph, your agent works like a surgeon operating without an X-ray, skilled but guessing at what's inside.

codeloom builds a queryable code graph from your entire codebase — extracting structure from 57 languages and formats, every function, class, import, call, and document — and exposes it to your AI agent. One install, and your agent stops grepping and starts understanding.

Quick Start

pip install codeloom

cd your-project/
codeloom install opencode    # for OpenCode
# or: codeloom install claude  # for Claude Code

Then tell your agent:

"Build a code graph for this project"

That's it. The graph auto-rebuilds when your session ends. No extra tokens, no extra commands, everything runs 100% locally.

---

What Changes

Before (grep)	After (codeloom)
Finds exact strings, misses semantic connections	Finds conceptually related code via vector + keyword + graph
Returns a flat list of file matches	Returns seeds plus a subgraph showing how they connect
No way to know what depends on what	codeloom impact "validate_token", finds all 47 callers instantly
Agent operates blind, guesses at relationships	Agent sees the full picture before making edits

Every search returns results like this:

seeds:
codeloom/core/pipeline.py:71
  │ def run_pipeline(source_dir: Path, ...) -> PipelineResult:
  │     """Run the full code graph build pipeline."""
storage/store.py:20
  │ class KnowledgeStore:

edges:
codeloom/core/pipeline.py:71 -calls-> storage/store.py:20
codeloom/core/pipeline.py:0 -defines-> codeloom/core/pipeline.py:71

Seeds tell you where relevant code lives. Edges tell you how it connects. Together they give your agent the full picture, no separate Read calls needed.

---

16 MCP Tools at a Glance

Three categories, one MCP server.

Search

Tool	What it does
search	5-signal HybridRAG, vector + keyword + graph + community fused into one ranking
search_keyword	FTS5 keyword-only (BM25), instant results for known names
search_vector	Semantic vector-only, finds conceptually similar code

Analysis

Tool	What it does
impact	Blast radius, every caller that depends on a symbol
dependencies	Upstream deps, what a symbol needs to function
context	360-degree view of a symbol, metadata, community, all edges, source snippet
detect_changes	Map unstaged git changes to affected graph nodes
explain_flow	Trace execution path through call chains
stats	Node/edge counts, kind distribution, god nodes
communities	Browse functional clusters (Leiden communities)
node	Details on a specific symbol with fuzzy name matching

Refactoring & Admin

Tool	What it does
rename	Find every location and reference for safe multi-file rename
export_subgraph	Export focused subgraph around a symbol as D3.js JSON
list_repos	List available code graphs with staleness status
build	Build or rebuild the code graph
watch	Watch files for changes and rebuild automatically

All tools are available via MCP (stdin/stdout), no HTTP server, no network, no configuration.

---

Languages & Formats

Structural extraction (functions, classes, calls, imports)

Full tree-sitter tags.scm-based resolution for 19+ core languages. All 58 languages get module-level indexing, source snippets, and embeddings — structural detail depends on optional tree-sitter-<lang> packages.

Ada	C	C#	C++	Common Lisp	Elixir	Fortran	Go
Groovy	Haskell	Java	JavaScript	Julia	Kotlin	Lua	Nix
Objective-C	OCaml	Perl	PHP	PowerShell	Python	R	Ruby
Rust	Scala	Shell	Solidity	Swift	Terraform	TypeScript	Zig
Assembly

Document & config extraction

CMake	CSV	CSS	DjVu	Dockerfile	DOCX
GraphQL	HCL	HTML	JSON	Make	Markdown
ODP	ODS	ODT	Org	PDF	RST
SQL	TOML	XLSX	XML	YAML

Plus 100+ natural languages for search queries via multilingual-e5-small embeddings. Search in any language, find results in any language.

---

AI Agent Integrations

One command per platform:

Agent	Install
Claude Code	codeloom install claude
OpenCode	codeloom install opencode
Codex CLI	codeloom install codex
Gemini CLI	codeloom install gemini
Cursor IDE	codeloom install cursor
Windsurf IDE	codeloom install windsurf
Cline	codeloom install cline
Aider CLI	codeloom install aider
Any MCP client	claude mcp add codeloom -- codeloom mcp

Each install writes context rules and registers hooks where supported. For OpenCode, it installs a skill, adds context rules to AGENTS.md, and registers the codeloom MCP server. Remove with codeloom uninstall <agent>.

---

Features

Search Before Grepping

5-signal HybridRAG fuses code vector search, text vector search, graph expansion, FTS5 keyword, and community signals into one ranked result set with subgraph edges. --kind, --file, and --include-tests filters narrow results without re-running.

Edit With Confidence

Run impact before editing to find every caller. Run context for a full symbol overview, community, all relationships, source snippet. Run detect_changes after edits to see which nodes are affected.

Auto-Rebuild

Stop/SessionEnd hooks detect changed files via git diff and trigger an incremental rebuild. Lock files prevent concurrent rebuilds. Zero manual intervention, the graph stays fresh after every session.

Incremental & Fast

SHA-256 content hashing skips unchanged files. Hot-start PageRank reuses previous importance scores. Parallel extraction (ProcessPoolExecutor) speeds up full builds by 24-64%. Typical incremental build: ~0.4s for no changes, ~4s for changes, 95%+ faster than a full rebuild. Model warmup (--warmup, default on) preloads embedding models on MCP server start so the first search is fast — disable with --no-warmup to save ~150MB RAM.

100% Local + MIT

No cloud services, no API keys, no telemetry. SQLite + FAISS for storage, sentence-transformers for embeddings. All data stays on your machine. MIT licence, no commercial restrictions, no licensing friction.

---

Performance

Benchmarks on a 2023 MacBook Pro (M2 Pro, 32GB RAM). All builds use parallel extraction (default: os.cpu_count() workers).

codeloom's own codebase (~3,500 lines, 90 files, 1,300 nodes)

Operation	Time
Full build	~14s
Incremental (changes)	~4s
Incremental (no changes)	~0.4s
Cold search (dual model)	~2.8s
Cold search (--fast)	~0.2s
Warm search	~0.08s
Cached search	<1ms

Synthetic stress tests (no embeddings)

Dataset	Files	Nodes	Build Time	Peak Memory
Tiny	10	119	0.7s	14 MB
Small	100	4,109	2.3s	16 MB
Medium	1,000	101,009	53.1s	393 MB
Large	5,000	205,009	164.9s	814 MB

Parallel extraction delivers 24-64% faster builds. Compact node storage (path interning, skipped empty attrs, RAM-free source snippets after persist) reduces peak memory by 10-22%. See docs/SCALING.md for detailed analysis.

• Embedding models: ~180MB, downloaded once to ~/.codeloom/models/
• Database: ~2MB (SQLite + FTS5 + FAISS indices)

---

Full CLI Reference

All commands output compact text by default (designed for AI agent consumption).

CLI Commands

Command	Description
build <dir>	Build code graph (--incremental, --git)
watch <dir>	Real-time file system monitor
search <query>	5-signal HybridRAG with subgraph + snippets
search-keyword <query>	FTS5 keyword matching only
search-vector <query>	Vector similarity only
search-graph <query>	Graph expansion only (BFS from vector seeds)
search-community <query>	Community cluster matching only
stats	Graph statistics
node <id>	Node details with fuzzy matching
communities	List or search communities
query	Interactive search REPL
export	Export as JSON, GraphML, or D3.js
visualize	Interactive HTML visualization
install [agent]	Install codeloom integration for AI agents
uninstall [agent]	Remove codeloom integration for AI agents
doctor	Check installation health
clean	Remove .codeloom/ database
mcp	Start MCP server
help [command]	Show categorised help with usage examples

MCP-Only Tools

These are available via codeloom mcp — see the MCP tools section above:

impact · dependencies · context · detect_changes · rename · explain_flow · export_subgraph · list_repos

---

Requirements

• Python 3.10+
• ~180MB disk for embedding models (cached on first use)

# Optional: PDF, DOCX, XLSX, ODF extraction
pip install codeloom[docs]

Development

pip install -e ".[dev]"
pytest
ruff check codeloom/

License

MIT License. See LICENSE for details.

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.