Cgraphy

cgraphy – Code Knowledge Graph MCP Server

[PyPI version] [Python versions] [License] [DOI]

Paper (DOI): https://doi.org/10.5281/zenodo.21422935
PyPI: https://pypi.org/project/cgraphy/
GitHub: https://github.com/pmgarg/cgraphy

cgraphy is a Python code knowledge graph and Model Context Protocol (MCP) server for AI coding agents such as Claude Code, Codex CLI, Cursor, and Gemini CLI.

cgraphy indexes any codebase into a knowledge graph — functions, classes and files as nodes; calls, imports, inheritance and git co-change history as edges — and serves compact, token-budgeted slices of it to AI assistants through the Model Context Protocol. Instead of re-reading dozens of files to orient itself on every prompt, an agent asks the graph and gets the relevant subgraph in a couple of thousand tokens.

• Any language. Full-fidelity extraction (calls, imports, inheritance) for Python, TypeScript/JavaScript, Java, Go, C, C++ and Rust; generic definition-level extraction for 20+ more via tree-sitter; config and docs files participate through summaries.
• Importance-ranked. PageRank over the code graph puts load-bearing symbols first in every answer.
• Token-budgeted. cgraphy_context expands the graph greedily around a symbol and stops exactly at your token budget — cost scales with the question, not the repo.
• Git-aware. --git-history mines commit history for files that change together (logical coupling), an edge type static analysis can't see.
• No API key. Semantic summaries are written by the host agent itself through the enrich loop; summaries survive re-indexing via content hashing.
• Zero infrastructure. One SQLite file in .cgraphy/. No services, no daemons, no vector database.
• Proven end-to-end. In 400+ controlled agent runs on SWE-bench Lite, the deployed configuration resolved 14 vs 8 of 57 real GitHub issues (official Docker harness). Indexes kubernetes (26K files, 219K nodes) in 49s, keeps it fresh in 1.6s cycles, answers queries in 1–152ms. All benchmarks and predictions are in this repo (Research paper and benchmarks).

Install

pip install cgraphy        # or: uv tool install cgraphy

Quick start

cd your-repocgraphy init          # one command: MCP config + agent steering + index

cgraphy init does three things:

1. Writes a project-scoped .mcp.json — picked up automatically by Claude Code in all its forms: CLI, VSCode extension, and the desktop app.
2. Appends a steering block to CLAUDE.md and AGENTS.md telling agents to consult the graph (cgraphy_overview → cgraphy_search → cgraphy_context) before reading files — this is what makes the graph actually replace bulk file reading. (Agents can't be forced, only steered: instruction files + persuasive tool descriptions + the tools being genuinely faster is the mechanism, and it works.)
3. Builds the index with git co-change history.

Or register the MCP server manually with your assistant:

Claude Code

claude mcp add cgraphy -- uvx cgraphy serve /path/to/repo

Codex CLI (~/.codex/config.toml)

[mcp_servers.cgraphy]command = "uvx"args = ["cgraphy", "serve", "/path/to/repo"]

Gemini CLI (~/.gemini/settings.json) / Cursor (.cursor/mcp.json)

{"mcpServers": {"cgraphy": {"command": "uvx",                            "args": ["cgraphy", "serve", "/path/to/repo"]}}}

The eight tools

Reading / orientation:

Editing / reviewing — the tools that make the graph part of the change loop:

Enrichment:

Retrieval is usage-aware: symbols an agent repeatedly asks about get a small, capped boost in future context expansion (telemetry stays in the local SQLite file; nothing leaves your machine).

Semantic search (optional)

pip install "cgraphy[semantic]"

Adds tiny static embeddings (model2vec, CPU-only, no torch) fused with FTS5 by reciprocal-rank fusion — closes the vocabulary gap between issue-style prose ("login broken") and code identifiers (validate_jwt).

The graph self-heals: tools detect stale files and re-index incrementally (changed files only) before answering.

Enriching the graph

Structure is extracted automatically; meaning comes from summaries. Tell your agent once:

enrich the cgraphy graph

It will loop cgraphy_enrich → cgraphy_store_summaries until every symbol has a one-line semantic summary. Summaries are keyed to a hash of each symbol's source, so editing one function invalidates only that summary.

For CI, cgraphy index --summarize pre-bakes summaries with your own Anthropic API key (pip install cgraphy[summarize], ANTHROPIC_API_KEY set).

Viewer

cgraphy view .        # http://localhost:8787

A dependency-free local page (bundled Cytoscape.js): search, color by kind, click for details, double-click to expand neighbors; co-change edges shown dashed.

Measuring the savings

python scripts/benchmark.py /path/to/repo "your question"

Prints the tokens an agent spends orienting via cgraphy (overview + search + context) versus reading every code file, and the reduction factor.

Localization benchmark (research harness)

python scripts/eval_localization.py /path/to/repo 50

Mines fix-like commits from the repo's history (subject = query, touched files = ground truth, co-change mining excludes evaluated commits), then scores an ablation ladder — FTS-only, +PageRank, +graph expansion, ±co-change edges — on hit@5/hit@10/MRR and token cost. No LLM calls, no human grading, fully reproducible. Results and a paper draft live in paper/.

How it works

1. cgraphy index walks the repo (respecting .gitignore + .cgraphyignore), parses each file with tree-sitter, and stores nodes and edges in .cgraphy/graph.db (SQLite + FTS5). Re-indexing is incremental by content hash.
2. A resolver links cross-file references (calls, imports, inheritance) by qualified name, best-effort; unresolved names are kept, never dropped.
3. PageRank runs over the edge graph; every query surfaces important symbols first. Search blends FTS5 relevance with rank.
4. cgraphy serve exposes the five MCP tools over stdio.
5. Optional: --git-history adds weighted co-change edges mined from git log.

Design details: docs/superpowers/specs/2026-07-08-cgraphy-design.md

Citing cgraphy

If you use cgraphy in research or tooling, please cite the paper:

@misc{garg2026cgraphy,  author = {Garg, Prateek Mohan},  title  = {Which Graph Signals Pay for Their Tokens? cgraphy: A Token-Budgeted            Code Knowledge Graph as a Portable Context Layer for AI Coding Agents},  year   = {2026},  doi    = {10.5281/zenodo.21422935},  url    = {https://doi.org/10.5281/zenodo.21422935},  note   = {Preprint}}

License

MIT

mcp-name: io.github.pmgarg/cgraphy