Codastre — code search & a live knowledge graph for your AI coding agents

Your coding agents guess at your codebase. Codastre makes them know it. It turns one or many git repositories into a topology-aware retrieval layer — dense semantic embeddings plus BM25 lexical search, fused with Reciprocal Rank Fusion — and a cross-repo knowledge graph. Any agent calls it over the Model Context Protocol (MCP).

Why teams use it

Fewer wasted tokens. Agents pull only the relevant lines instead of whole files or naive grep dumps — lower API cost per task.
Fewer wrong-file edits. Branch-aware indexing keeps context current, so agents stop editing stale or unrelated code.
Your source never leaves your machine. Queries return masked paths, line ranges, and scores — never code. Paths are HMAC-masked with a key held in your OS keychain.
Cross-repo awareness. Trace Kafka producers/consumers, HTTP and gRPC calls, and shared packages across repository boundaries.

Deploy it your way

Run Codastre self-hosted (open source, Helm chart or Docker Compose) so your code and index never leave your network, or use the managed cloud with per-tenant isolation and SSO.

MCP-native

Four tools your agent already understands — REGISTER, SYNC, QUERY, and GRAPH — over streamable HTTP. Works with Claude Code, Codex, opencode, and any MCP client.

Create a workspace · Install the CLI · About · Privacy