CodeCanvas MCP

Evidence-backed code intelligence for Python coding agents.

CodeCanvas is a local static-analysis Model Context Protocol server. It answers questions about call paths, control flow, and change impact without forcing an agent to grep through an entire repository and infer the architecture from partial text matches.

On a blinded three-task Google ADK holdout, the logic_flow-only profile used 52.58% fewer server-reported input + output tokens than a built-in-tools baseline, with mean rubric scores of 99.5/100 versus 100/100. Uncached input + output fell 14.39%; the repository documents the full methodology and the simple task that regressed.

Quick start

Python 3.10 or newer is required.

claude mcp add codecanvas -- uvx codecanvas-mcp

For another MCP client, use the equivalent stdio configuration:

{
  "mcpServers": {
    "codecanvas": {
      "command": "uvx",
      "args": ["codecanvas-mcp"]
    }
  }
}

Pass an absolute project_path on the first tool call. CodeCanvas remembers the last explicitly selected project for the rest of the server session.

Tools

Evidence-first responses

Static analysis is not runtime truth, so CodeCanvas exposes uncertainty instead of hiding it. Successful responses identify the selected analysis_root and report evidence strength, inferred or ambiguous edges, truncation, and whether the result is safe to summarize as an unconditional claim.

verify_claim returns true, false, or uncertain. Unsupported qualifiers and inferred-only paths cannot silently become a definite true.

Scope and safety

• CodeCanvas analyzes Python source with AST-based call-graph and control-flow analysis.
• FastAPI routes and Depends() chains, scripts, and distributed package exports are recognized as entrypoints.
• Compatible analysis is cached locally under <project>/.codecanvas/.
• Dynamic imports, reflection, monkey patching, and runtime-only values may remain unresolved and are reported as qualifications where possible.
• simulate_state_transition imports and executes trusted project code in a separate process. It is not a security sandbox; the imported code may access the filesystem, network, or subprocesses and may have import-time effects.

The complete guide, agent instruction snippet, benchmark command, and development setup are available in the GitHub repository.

License

CodeCanvas MCP is open-source software licensed under the MIT License.