Graphkeeper

GraphKeeper

Install • Quickstart • CLI Reference • Comparison • FAQ

A local-only CLI that mines your git log for which files actually change together, then hands an AI coding agent a queryable answer instead of a grep across the whole history.

npx graphkeeper-cli build
npx graphkeeper-cli query co-change src/git.ts

Files that historically change alongside "src/git.ts":

     1  src/store.ts
     1  src/types.ts
     1  test/git.test.ts
     1  test/store.test.ts
     1  test/test-helpers.ts

(Real output from running GraphKeeper against its own repo, this early in its history. Co-change counts grow as a codebase accumulates more commits.)

No server, no account, no embeddings API. Every byte of output comes from git log on the repo you already have checked out.

Install

GraphKeeper ships two independent, equally first-class packages. Pick whichever fits your toolchain, or install both. Both mine the same git log co-change signal and share one on-disk .graphkeeper/graph.json schema, so a store built by either can be read back by the other.

# npm -- JavaScript/TypeScript CLI + library
npm install -g graphkeeper-cli
# or run it once with no install
npx graphkeeper-cli build

# PyPI -- Python CLI + library (genuine port, not a wrapper around the Node binary)
pip install graphkeeper-cli

The npm package requires Node.js 18 or later; the Python package requires Python 3.9 or later. Both require git on your PATH. The Python package's CLI entry point is also graphkeeper (e.g. graphkeeper build). See python/README.md for the Python-specific walkthrough, and CHANGELOG.md for each distribution's version history.

Quickstart

Run it against any git repo, including this one:

git clone https://github.com/RudrenduPaul/GraphKeeper.git
cd GraphKeeper
graphkeeper build

GraphKeeper build complete: /path/to/GraphKeeper

Co-change graph: 4 commit(s) analyzed, 80 file pair(s) found
graphify enrichment: skipped -- graphify was not found on PATH. Install it with
`uv tool install graphifyy` (or `pipx install graphifyy`) for symbol/call-graph
enrichment; GraphKeeper works fine without it, in co-change-only mode.

Wrote /path/to/GraphKeeper/.graphkeeper/graph.json

Now query it:

graphkeeper query co-change src/git.ts

Files that historically change alongside "src/git.ts":

     1  src/store.ts
     1  src/types.ts
     1  test/git.test.ts
     1  test/store.test.ts
     1  test/test-helpers.ts

If graphify is installed (uv tool install graphifyy), graphkeeper build automatically shells out to its local, no-API-key graphify extract --code-only and merges its symbol/call-graph into the same store, unlocking call-graph queries:

graphkeeper query calls mineCoChange

mineCoChange() (src/git.ts)

Calls (2):
  --> assertIsGitRepo()
  --> runGit()

Called by (1):
  <-- build()

(Also real output, from running graphkeeper build against this repo with graphify installed.)

Without graphify installed, that same command explains exactly why the answer isn't available instead of crashing or returning an empty result:

Call-graph query for "mineCoChange" is not available.

graphify was not found on PATH. Install it with `uv tool install graphifyy`
(or `pipx install graphifyy`) for symbol/call-graph enrichment; GraphKeeper
works fine without it, in co-change-only mode.

Every command also supports --json for scripts and agents:

graphkeeper query co-change src/git.ts --json

{
  "file": "src/git.ts",
  "results": [
    { "file": "src/store.ts", "count": 1 },
    { "file": "src/types.ts", "count": 1 },
    { "file": "test/git.test.ts", "count": 1 },
    { "file": "test/store.test.ts", "count": 1 },
    { "file": "test/test-helpers.ts", "count": 1 }
  ]
}

Features

• Mines real commit history, not a static snapshot. graphkeeper build runs git log --no-merges --name-only across the full history of the repo and counts every file pair that changed together in the same commit. There's no guessing at coupling from folder structure or import statements alone; the answer comes from how the codebase actually got edited over time.
• --max-files-per-commit protects the signal. A single vendoring commit or mass reformat that touches 400 files would otherwise pollute every pair in that commit. The default cap (100 files) skips commits above that threshold so real coupling doesn't drown in noise.
• Optional call-graph enrichment, never required. When graphify is on PATH, graphkeeper build shells out to its local graphify extract --code-only --no-cluster and merges the resulting symbol/call edges into the same store, unlocking graphkeeper query calls. Without graphify, GraphKeeper still works in co-change-only mode and says so plainly instead of failing.
• Every command has a --json mode. graphkeeper query co-change <file> --json and the equivalents return machine-readable output, so an agent's calling code parses a real data structure instead of scraping text.
• Two from-scratch implementations, one schema. The npm package (src/, TypeScript) and the PyPI package (python/src/graphkeeper/, Python) are independent ports, not a wrapper of one around the other. Both read and write the same .graphkeeper/graph.json, so a store built with one CLI is queryable from the other.
• Every subprocess call uses an argv array, never a shell string. Git and graphify are both invoked through spawnSync/subprocess.run with a list of arguments, so a crafted commit message or filename in the repo being analyzed can't be interpreted as shell syntax.

CLI Reference

Usage: graphkeeper [options] [command]

Options:
  -V, --version           output the version number
  -h, --help              display help for command

Commands:
  build [options] [path]  Mine git history for co-change and (if available)
                          merge in graphify's symbol/call graph
  query                   Query the GraphKeeper store built by
                          `graphkeeper build`
  help [command]          display help for command

graphkeeper build [path]

Walks path (default: current directory), runs git log --no-merges --name-only across the whole history, and counts how often each pair of files was touched in the same commit. Writes the result to .graphkeeper/graph.json.

If graphify is detected on PATH, build also runs graphify extract <path> --code-only --no-cluster (graphify's own headless, local, no-API-key AST extraction path) into a directory inside .graphkeeper/, and merges its nodes/edges into the same store. The build output always states plainly whether that enrichment was included, and why it was skipped if not.

graphkeeper query co-change <file>

Lists files that historically changed alongside <file>, ranked by how many commits touched both.

Exit code 0 when results are found, 1 when there's no co-change data for that file yet, 2 on a usage or filesystem error.

graphkeeper query calls <symbol>

Shows callers and callees of <symbol>, using graphify's calls edges from the most recent build. Only meaningful when that build included graphify enrichment; if it didn't, this prints a clear explanation of why instead of a crash or a silent empty result.

Exit code 0 when the symbol is found, 1 when it isn't (or enrichment wasn't available), 2 on a usage or filesystem error.

MCP Server

GraphKeeper ships a Model Context Protocol server so an AI coding agent (Claude, Cursor, or any MCP-compatible client) can mine co-change history and query the graph directly, without a human invoking the CLI by hand.

Install the extra:

pip install "graphkeeper-cli[mcp]"

Add it to your MCP client's config (for Claude Desktop, claude_desktop_config.json):

{
  "mcpServers": {
    "graphkeeper": {
      "command": "uvx",
      "args": ["--from", "graphkeeper-cli", "graphkeeper-mcp"]
    }
  }
}

The server exposes one tool, run, that shells out to the published graphkeeper-cli npm package (via npx graphkeeper-cli) with the given subcommand and arguments plus --json, and returns the parsed JSON result:

run(["build", "."])
run(["query", "co-change", "src/git.ts"])

Transport is stdio, so there is nothing to host: the MCP client spawns the server as a local subprocess. Source: python/src/graphkeeper/mcp_server.py.

Library API Reference

Both packages expose a real importable surface, listed below. The command line is a thin wrapper over the same functions.

TypeScript / JavaScript (graphkeeper-cli on npm)

import { build, queryCoChange } from "graphkeeper-cli";

const result = build(".");
console.log(`${result.store.commitsAnalyzed} commit(s) analyzed`);

const coChange = queryCoChange(result.store, "src/git.ts");
for (const row of coChange.results) {
  console.log(row.count, row.file);
}

Types (GraphKeeperStore, CoChangeEdge, GraphifyNode, GraphifyEdge, GraphifyRawGraph, GraphifyEnrichment, BuildOptions, BuildResult, CoChangeQueryResult, CallsQueryResult) ship as .d.ts declarations alongside the compiled output; no separate generated docs site exists yet.

Python (graphkeeper-cli on PyPI)

from graphkeeper import build, query_co_change

result = build(".")
print(f"{result.store.commits_analyzed} commit(s) analyzed")

co_change = query_co_change(result.store, "src/git.py")
for row in co_change.results:
    print(row.count, row.file)

graphkeeper.__all__ exports: build, query_co_change, query_calls, find_graphify_node, normalize_file_arg, detect_graphify, run_graphify_enrichment, mine_co_change, unquote_git_path, GitError, resolve_repo_root, read_store, write_store, graph_file_path, PathSafetyError, plus the GraphKeeperStore, CoChangeEdge, GraphifyNode, GraphifyEdge, GraphifyEnrichment, BuildOptions, BuildResult, CoChangeQueryResult, CoChangeResultRow, and CallsQueryResult dataclasses. Same names as the TypeScript export table above, in snake_case. No separate generated docs site exists yet; the docstring at the top of python/src/graphkeeper/__init__.py covers the same ground as this section.

Comparison

("Not publicly documented" is used instead of a guess anywhere a competitor doesn't publish the number. Verified against each project's own README, pricing page, or public site as of August 2026.)

What Is GraphKeeper, and Why Does It Exist

GraphKeeper is a local CLI and library, published to both npm and PyPI as graphkeeper-cli, that mines git log for file-level co-change: which files have historically been edited in the same commit as a given file. It writes that data to a single JSON file, .graphkeeper/graph.json, and answers queries against it with no network calls.

It exists because an AI coding agent working solo on a codebase it doesn't already know well has no fast way to answer "what else usually changes when I touch this file?" without running its own git log --name-only scan and tallying the results by hand, every time it's asked. GraphKeeper precomputes that answer once and makes it queryable, including in a --json form a script or agent can parse directly.

GraphKeeper does not reimplement symbol or call-graph extraction. graphify (100K+ GitHub stars, Apache-2.0, pip install graphifyy) already does that across 36 tree-sitter grammars and ships as a slash-command skill for Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, and 15+ more assistants (20+ total). When graphify is on PATH, graphkeeper build shells out to its local graphify extract --code-only --no-cluster and merges the result into the same store, unlocking graphkeeper query calls. Without graphify, GraphKeeper still works, in co-change-only mode, and says so directly instead of failing.

How It Works

1. graphkeeper build runs git log --no-merges --name-only (via a safe argv-array subprocess call, never a shell string) across the whole repo history.
2. For every commit, it counts every pair of files that changed together. Commits touching more than --max-files-per-commit files (default 100) are skipped, so a single vendoring or mass-reformat commit can't drown out real signal.
3. If graphify is detected on PATH, GraphKeeper also runs graphify extract <path> --code-only --no-cluster, graphify's own local, no-LLM, no-API-key extraction mode, into a directory inside .graphkeeper/, then merges its nodes/edges into the same store.
4. The merged result is written once, atomically, to .graphkeeper/graph.json.
5. graphkeeper query reads that file back and answers co-change or call-graph questions against it. No network calls, ever.

Security

• Every git and graphify invocation uses an argv array passed directly to the OS (spawnSync in TypeScript, subprocess.run with a list in Python), never a shell string, so commit messages, file names, or repo paths can't be interpreted as shell syntax.
• .graphkeeper/ output paths are checked against the resolved repo root before every write (symlinks included, via fs.realpathSync), so a maliciously crafted repo can't redirect GraphKeeper's writes outside .graphkeeper/.
• No telemetry, no network calls, no secrets. The only files GraphKeeper reads are git log output and, optionally, graphify's own graph.json; the only file it writes is .graphkeeper/graph.json.

See SECURITY.md for the vulnerability reporting process.

FAQ

Is GraphKeeper a general codebase knowledge-graph indexer?

Not on its own. graphkeeper build mines git log for file-level co-change and writes those edges to .graphkeeper/graph.json. That file only becomes a symbol/call graph too if graphify is installed and gets merged in during the same build. Without graphify on PATH, the store holds co-change data only, and graphkeeper query calls says so directly instead of returning an empty result.

What does GraphKeeper actually give an agent that grep or git log don't?

A pre-computed, queryable answer to "which files change together here," so an agent doesn't have to run its own git log --name-only scan and tally the results by hand on every question. --json on every command makes that answer script-consumable rather than something a human has to read and re-type.

How do I install it, and does it work on Windows?

npm install -g graphkeeper-cli (Node.js 18+) or pip install graphkeeper-cli (Python 3.9+); both need git on PATH. Neither package contains OS-specific branches or native bindings, and the PyPI listing is classified Operating System :: OS Independent, so it runs the same way on Windows, macOS, and Linux anywhere git and a supported Node or Python runtime are available.

How is this different from graphify, the tool it links to for enrichment?

They answer different questions. graphify extracts symbols, imports, and call graphs straight from source via tree-sitter, across 36 languages; GraphKeeper mines commit history for which files were historically edited together, a signal graphify has no reason to compute. GraphKeeper shells out to graphify's own local extract command when it's present and merges the result in, rather than reimplementing tree-sitter parsing from scratch. Neither replaces the other; see the Comparison table above for how GitNexus, Greptile, and Augment Code differ from both.

What actually breaks GraphKeeper, or gives an empty result?

Two real cases, both documented, neither a crash: a shallow git clone (GitHub Actions' default fetch-depth: 1) has no history to mine, so build reports 0 commit(s) analyzed and writes an empty co-change graph; full history (fetch-depth: 0) is required. Separately, query calls only returns results if the most recent build ran with graphify on PATH; if it didn't, the command explains that plainly (graphify was not found on PATH...) instead of pretending the symbol doesn't exist.

Is it safe to run against a repo I don't fully trust?

Every git and graphify call goes through an argv array straight to the OS (spawnSync / Python's subprocess.run with a list, never a shell string), so filenames or commit messages can't be interpreted as shell syntax. Every .graphkeeper/ write is checked against the resolved repo root, symlinks included, before it happens. There are no network calls anywhere in the tool, so nothing about the repo you point it at leaves your machine.

Is the npm CLI just a wrapper around the Python one, or vice versa?

Neither. They're two independent, from-scratch implementations (src/ for TypeScript, python/src/graphkeeper/ for Python) that happen to agree on the same .graphkeeper/graph.json schema, the same subcommands, flags, and exit codes. A store built by one can be read by the other. The Python port's own test suite (ported from the TypeScript vitest suite) is 78 tests, run against a real subprocess CLI invocation, not a mock of the other language's output. Both suites pass in a clean install as of this writing: 78/78 on the TypeScript side (npm test), 78/78 on the Python side (pytest).

What license is this under, and can I use it commercially?

Apache License 2.0, for both the npm and PyPI packages, with no dual licensing and no separate commercial tier. That permits commercial use, modification, and redistribution, with attribution and the standard Apache patent grant. See LICENSE for the full text.

Contributing

Issues and PRs welcome. To build the TypeScript package from source:

git clone https://github.com/RudrenduPaul/GraphKeeper.git
cd GraphKeeper
npm install
npm run build
npm test
npm run lint
npm run typecheck

For the Python package, see python/README.md. Full contribution guidelines covering both codebases are in CONTRIBUTING.md.

License

Apache License 2.0