Skip to content

Overview

A persistent, incremental knowledge graph of a codebase, exposed to an AI agent for token-efficient, impact-aware code review. Ask “who calls this?”, “what does this change break?”, and “review my diff” without pasting whole files into the prompt.

  • Graph: parse source into a call/import/inheritance graph, update incrementally, embed nodes for semantic search, and export/import a portable graph artifact (build on CI, import on a laptop).
  • Query: predefined relationship queries (callers, callees, imports, tests, inheritors), name/keyword/vector search, and blast-radius impact analysis from changed files.
  • Review: auto-detect changed files from a git diff and return a structural summary, impacted nodes, and source snippets as review context.
  • Security: scan graph nodes with a heuristic or Semgrep engine and emit JSON or SARIF reports.

Domain tools graph, query, review, and security, each driven by an action parameter, plus the universal config (status, runtime settings, credential setup) and help; config__open_relay reopens the browser credential form. Every action and parameter is listed in the tools reference.

Defaults to stdio and runs against a local checkout; it ships Docker :http builds for advanced self-host but has no hosted remote-relay mode by default. Works with any stdio MCP client – Claude Code, Codex, Gemini CLI, Cursor, and Windsurf. See the modes overview.

The graph, heuristic security scan, and most queries work with no credentials. Optional cloud embeddings and LLM node summaries are unlocked with your own keys and model chains, all supplied as environment variables:

  • EMBEDDING_MODELS, SUMMARY_MODELS – optional CSV model chains for node embeddings and function summaries; empty falls back to the bundled local model or disables the feature.
  • Provider API keys (JINA_AI_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY, COHERE_API_KEY).

Use placeholders in shared config and never commit real keys. Full walkthrough in setup.