P-020 · Case study
OPEN SOURCECitegraph
Open-source agent system that answers questions about a Python codebase with path:line citations. The server renders each citation from retrieved evidence, so the model can't invent one.
PythonMCPFastAPINeo4jAnthropic API (Claude, prompt caching)PydanticDockerpytestmypyGitHub Actions
P-020 / Register plateOPEN SOURCE
8,340
Nodes in the FastAPI code graph
PythonMCPNeo4j
More detail
Citegraph parses a repository into a Neo4j code graph and plans each question as a typed, validated set of allowlisted tool calls, run in bounded dependency waves. Five MCP servers split planning, indexing, graph retrieval, code analysis and memory behind a FastAPI gateway with HTTP, SSE and WebSocket APIs. It indexes FastAPI 0.162.0 out of the box: 537 files, 8,340 nodes and 12,067 relationships. Ten sample questions ran live on Claude Opus 5.5 with every plan accepted and no degraded answers.