minga-graph

Turn a codebase (code, docs, SQL schemas, dependency manifests) into a local knowledge graph you can query instead of grepping. Every edge says how it was found and points to the exact file:line. Built for Claude Code; works from any terminal and through MCP.

Русская версия →

Why another one

minga-graph is inspired by the idea behind graphify and written from scratch with different priorities: nothing leaves your machine and nothing outside your control runs.

Honest trade-off: fewer languages for now (see below). Adding languages is on the roadmap.

Related MCP server: koragraph

Install

Needs Python 3.10+.

pipx install git+https://github.com/dmminga1983/minga-graph
minga-graph install          # adds the /minga-graph skill to Claude Code (~/.claude/skills)

Then in Claude Code: /minga-graph .

Without installing: python -m minga_graph build . from a clone.

Use

minga-graph build .                                # incremental: only changed files are re-parsed
minga-graph query "how does the client reach the users table"
minga-graph explain ApiClient
minga-graph path Profile formatUser
minga-graph impact formatUser                      # what breaks if I change it
minga-graph diff                                   # structure changes since the previous build
minga-graph cycles
minga-graph secrets .                              # exit 1 if a hard-coded secret is found

Output in minga-out/:

graph.html        interactive graph, opens offline in any browser
GRAPH_REPORT.md   key nodes, subsystems, surprising links, cycles, unused candidates, secrets
graph.json        the full graph for scripts and the MCP server

Example:

$ minga-graph path Profile formatUser
Shortest path (4 hops):
  Profile --calls--> loadProfile --calls--> getUser --calls--> normalize --calls--> formatUser

Evidence:
  Profile --calls--> loadProfile [EXTRACTED] bound by import from ../api/client @ src/ui/Profile.jsx:6
  loadProfile --calls--> getUser [INFERRED] unique method name (receiver type unknown) @ src/api/client.ts:30
  getUser --calls--> normalize [EXTRACTED] this.normalize inside its class @ src/api/client.ts:14
  normalize --calls--> formatUser [EXTRACTED] bound by import from ../util/format @ src/api/client.ts:18

Exit codes: 0 ok, 1 error or secrets found, 3 ambiguous name (candidates are printed), 4 not found.

What it understands

Ignored by default: .gitignore rules (nested too), .mingaignore, node_modules, virtualenvs, build output, lock files, and all .env files (never read).

Edge confidence

• EXTRACTED: the source binds the name explicitly: same-file definition, an import binding, self.x/this.x in the defining class, SQL REFERENCES, a Markdown link.

• INFERRED: matched by a unique project-wide name (for example a method called on an object whose type is unknown). Open the @ file:line before relying on it for a risky change.

• Ambiguous or very generic names (get, run, update…) are dropped. No edge is better than a wrong edge.

MCP server

claude mcp add minga-graph -- minga-graph mcp /path/to/project

Tools: graph_query, graph_explain, graph_path, graph_impact, graph_stats. Read-only, reloads the graph when it is rebuilt.

Keep the graph fresh

minga-graph hook install      # git post-commit hook, rebuilds in the background

Options

--lang ru Russian report · --exclude PATTERN extra ignore rule · --resolution 1.5 smaller communities · --node-limit N nodes drawn in HTML · --full ignore cache · --json machine-readable summary.

Owner control

This repository includes an owner-control kit (CODEOWNERS, an importable branch ruleset, safe SHA-pinned CI, Dependabot, private vulnerability reporting). See docs/OWNER_CONTROL.md for what it does and the 3-minute one-time setup. The kit can be copied into any repository.

Roadmap

Go, Java, C#, PHP and Rust extractors; type-aware receiver resolution for obj.method(); cross-repo graphs.

License

MIT. Not affiliated with Graphify Labs; this is an independent implementation.