// the find
DeusData/codebase-memory-mcp
High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.
An MCP server that indexes codebases into a SQLite-backed knowledge graph using tree-sitter AST parsing, then exposes structural queries (call graphs, dead code, impact analysis) to your AI coding agent instead of having it grep through files. Ships as a single static binary with 158 vendored grammars and a bundled embedding model for semantic search. Aimed at developers who spend too many tokens on file-by-file exploration in Claude Code, Cursor, and similar tools.
The zero-dependency binary is the right call — 158 tree-sitter grammars, a 768d int8 embedding model, and all LSP type-resolution logic compiled in together means nothing to install, nothing to break. The RAM-first indexing pipeline (LZ4 compressed reads, in-memory SQLite, single flush at end) is genuinely engineered: Linux kernel in 3 minutes for 28M LOC is not a marketing number, that's just what you get when you avoid disk I/O in the hot path. The team-shared graph artifact (`.codebase-memory/graph.db.zst` committed to the repo) is a practical feature that other similar tools miss — teammates skip the full reindex on clone. Supply chain hygiene is unusually serious: SLSA Level 3, OpenSSF Scorecard, VirusTotal on every release binary, HMAC-signed artifacts. Most tools at this star count don't bother.
The Hybrid LSP — reimplementing type resolution for 11 languages from scratch in C — is the part most likely to quietly produce wrong answers. It works on the happy path, but complex Rust trait dispatch, Python dynamic metaclasses, or TypeScript conditional types will fall through the cracks in ways the graph won't tell you about. The daemon coordination model has a sharp edge: all active CBM processes must be the exact same version, same binary build, same cache root. Upgrading mid-session on a machine with multiple projects open fails with an admission conflict. The token-reduction benchmarks (99%, 120x) come from the authors' own arXiv preprint across 31 repos — independently plausible but not independently verified, and the gains collapse for tasks that genuinely need to read source rather than traverse structure. The 43 supported client configurations table reads like a feature count, but it's operational surface area: each client integration has slightly different hook semantics, and the README's caveats about what's withheld where (Cursor hooks have a race, Cline doesn't consume hook output, Windows gets different behavior than macOS) will matter when something silently doesn't work.