code-graph indexes a codebase on your machine and answers the questions a change needs: where is the code for this request, what calls it, which tests reach it, and what did my change touch. It answers from a local SQLite file without calling a model, so your agent reads a few hundred lines it can verify instead of opening files one by one.
| If you want to… | Read |
|---|---|
| Install it and ask a first question | Getting started |
| Understand how the answers are built and how far to trust them | Concepts |
| Look up a command or an MCP tool | CLI reference · MCP reference |
| Task | Guide | Main commands |
|---|---|---|
| Find where to change for a request, an issue or a stack trace | Find the code | locate, search, find |
| Know what breaks and which tests to run before you change | Impact and tests | impact, verify, callers |
| Check what a change actually touched | Review a change | review, check |
| Give the same answers to Codex, Claude Code or any MCP client | Agents and MCP | skill, mcp |
| Browse a project visually | Dashboard | explore |
| Start from Payment or another functional purpose, then open the code and tests | Two maps | flows, prepare, verify |
| Hand a change to one model and have the other review it | Change with an agent | change |
| Share an offline picture of the call graph | Export | export |
| Work on large repositories and keep the index fresh | Large projects | index, .cgignore, stats |
- CLI reference: every command and option, generated from the code.
- MCP reference: every tool and parameter, generated from the code.
- Formats: the JSON answers of
locate,search,impactandverify. - Languages: what is resolved in each language, and where the analysis abstains.
- Benchmarks: measured indexing speed, retrieval and agent pilots, with their limits.
- FAQ and Troubleshooting.
code-graph pays its way when the code you need is not in a file you already have open: a shared helper with many callers, a request worded differently from the code, a project too large to read. On a few small files, reading them is as cheap and the graph adds nothing; the benchmarks show this too.
It is a static analysis. A call that is only decided at run time (dynamic dispatch, reflection, dependency injection, routes registered by string) can be missing from the graph, and every answer says how a call was resolved. Treat the answers as a fast, verifiable guide to the code, and let the tests decide.
Every example in this guide runs on example/shop, a small order service with tests,
and shows real output: tools/docs_examples.py re-runs the commands and a test fails when a page
no longer matches what the code prints.