CodeGraph (Rust) MCP

Install

Works in

• Claude Code
• Cursor
• Codex
• SourceCraft
• Koda
• Coddy

Manual install

git clone https://github.com/Jakedismo/codegraph-rust
cd codegraph-rust
./install-codegraph-full-features.sh

What it does

CodeGraph for Rust parses a project's sources into a syntax tree, links symbols into a graph and stores it with vector embeddings in SurrealDB, which is built into the `codegraph` binary. Each project gets its own store in `.codegraph/db`. The indexing tier sets how deep the parsing goes: `fast` takes nodes and core edges, `balanced` adds language-server symbols and docs, `full` turns on every analyzer, dataflow and architecture. The agent connects the server over MCP and gets four tools. `agentic_context` gathers the context you need and answers questions about the code, `agentic_impact` shows dependency and call chains, that is, what breaks when you change something, `agentic_architecture` describes structure and public APIs, `agentic_quality` finds complex and tightly coupled spots. Inside each tool a model from `.env` works: it picks graph operations itself, chains them and returns file paths, line numbers and short snippets. The same tools are available from the shell with `codegraph agent context "question"`. `codegraph init` offers project-local hooks for Claude Code and Codex, appends a managed section to `AGENTS.md` and `CLAUDE.md` and starts indexing. The `--watch` flag refreshes the index while the server runs.

Who it is for. For developers who work in large repositories with Claude Code, Cursor or Codex and want the agent to look at the links in the code first instead of reading files one by one.

Example request

Build a code graph of this project and show which modules break if I change the signature of `UserService`

MCP

Transport

stdio

Authentication

not required

Limitations

• The author did not specify a license, so rights to use and copy are undefined, check with the author. The project text names MIT, but there is no `LICENSE` file in the repository.
• The embedded store is opened by one process per project. Stop a running `codegraph start` before `codegraph index`, and connect one client per project. A separate daemon cannot share the store with a running server; a shared database needs your own SurrealDB server and the `CODEGRAPH_SURREALDB_URL` variable.
• The `balanced` and `full` tiers need language servers: `rust-analyzer`, `typescript-language-server`, `pyright-langserver`, `gopls`, `jdtls` or `clangd`. If the one you need is missing, indexing fails at once; for Rust `rust-analyzer --version` is checked from the project folder.
• The `fast` tier is the default and turns off language servers, enrichment and dataflow, so `Uses` and `References` edges stay out of the graph. How complete the answers are depends on the chosen tier and the model.
• Analysis across several repositories and custom graph schemas are on the author's roadmap and do not exist yet.

How to disable. Remove the `codegraph` block from the agent's MCP config and stop `codegraph start`. Delete a project's index with `rm -rf <project>/.codegraph/db`. The Claude Code and Codex hooks added by `codegraph init` live in the project settings, and the `# codegraph` section in `AGENTS.md` and `CLAUDE.md` is removed by hand.

README in short

The repository holds a Cargo workspace of separate crates, four install scripts (full, cloud, fast for macOS and basic), a sample config in `config/example.toml`, installation and usage guides, a description of the agent command line and the `agentic_test_cases.py` test scenarios. The main project page shows a sample tool answer, the table of indexing tiers, accuracy measurements per tier and a diagram of the MCP client, server and graph store.