DeusData/codebase-memory-mcp

A single-binary MCP server, written in C, that indexes codebases into a persistent knowledge graph for fast, token-efficient structural queries by AI coding agents.

What it is

codebase-memory-mcp parses a codebase with tree-sitter and builds a persistent, SQLite-backed knowledge graph of functions, classes, call chains, HTTP routes, and cross-service links. It is a structural analysis backend, not an LLM: it answers structural queries and relies on the connected MCP client to be the intelligence layer. It targets developers using MCP-compatible coding agents who want graph queries in place of repeated file-by-file grep and read cycles. It ships as a single static binary for macOS, Linux, and Windows with no runtime dependencies.

Key features

Indexes 158 languages using vendored tree-sitter grammars compiled into the binary.
Adds Hybrid LSP semantic type resolution for twelve languages: Python, TypeScript/JavaScript/JSX/TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl.
Exposes 15 MCP tools spanning search, call-path tracing, architecture overview, git-diff impact mapping, Cypher-like queries, dead-code detection, cross-service HTTP linking, and ADR management.
Semantic search using bundled Nomic nomic-embed-code embeddings compiled into the binary, plus BM25 full-text search via SQLite FTS5 and structural regex search.
Cross-service detection (HTTP, gRPC, GraphQL, tRPC, pub-sub channels), cross-repo linking, and infrastructure-as-code indexing for Dockerfiles, Kubernetes manifests, and Kustomize overlays.
Optional 3D graph visualization UI at localhost:9749, and a team-shared graph artifact committed as a single compressed file so teammates can skip reindexing.

Tech stack

A pure C core with zero dependencies, statically linked, combining tree-sitter grammars, Aho-Corasick matching, LZ4 compression, in-memory SQLite, and FTS5.
Bundled Nomic nomic-embed-code embeddings (768-dimension int8) compiled into the binary, so semantic search needs no API key, Ollama, or Docker.
An optional graph UI built with TypeScript, React, and Vite under graph-ui/, shipped as a separate ui build variant.
Persistence to ~/.cache/codebase-memory-mcp/ via SQLite; the team artifact is zstd-compressed.
Distributed through npm, PyPI, Homebrew, Scoop, Winget, Chocolatey, AUR, and go install; building from source needs a C/C++ compiler and zlib.
Licensed MIT.

When to reach for it

Giving an MCP coding agent fast structural understanding of a large codebase.
Reducing token usage relative to file-by-file exploration; the README cites roughly 99% fewer tokens across five structural queries.
Impact analysis on uncommitted changes, dead-code detection, call-path tracing, and one-call architecture overviews.
Mapping cross-service and cross-repo architecture across an indexed fleet.
Sharing a prebuilt graph artifact with teammates to avoid a full reindex.

When not to reach for it

When you expect a built-in LLM; this is a structural backend only and depends on your MCP client for natural-language reasoning.
Workflows without an MCP-capable client, though a CLI mode exists for direct queries.
Environments that cannot run an unsigned binary which also writes to agent configuration files; the README recommends auditing the signed, checksummed releases first.
Questions about runtime behavior or intent that go beyond code structure.

Maturity signal

The project reached over 33,000 stars since its creation in February 2026, was last pushed in July 2026, and is MIT-licensed, which points to very active and fast-moving development with rapid adoption. The high open-issue count (289) is consistent with a young project absorbing heavy usage. Releases are signed, checksummed, and scanned by multiple antivirus engines, and the design is described in an arXiv preprint (2603.27277).

Alternatives

graphify — use it if you prefer that workflow; the README explicitly compares its team artifact to graphify's graphify-out/ directory.
Single-language language servers (tsserver, pyright, gopls, rust-analyzer, and similar) — use them when you want full IDE-grade semantic analysis for one language rather than a cross-language graph built for agents; the README cites these as compatibility inspirations.
Plain grep or ripgrep with file reads — use them for small codebases where the indexing overhead is not worth it, which is the baseline the README benchmarks against.

Notes

The entire engine is pure C with zero dependencies and ships as one static binary.
Both the 158 tree-sitter grammars and the embedding model are compiled into the binary, so there is no external model download.
There is deliberately no built-in LLM; the reasoning is that the MCP client already acts as the query translator.
The team-shared artifact (.codebase-memory/graph.db.zst) auto-creates a merge=ours gitattributes line to avoid binary merge conflicts.
Processing is entirely local with no telemetry, so performance and memory issues must be captured and shared manually through the documented diagnostics files.

Wiki: DeusData/codebase-memory-mcp