Codebase Memory MCP

MIT

code-intelligenceknowledge-graphtree-sitterlocal-first

Summary

Local code-intelligence MCP server that indexes a repository into a persistent knowledge graph, so structural questions cost one query instead of a file-reading spree.

Features

Persistent code knowledge graph queried with Cypher-like syntax
Tree-sitter parsing across 158 languages, hybrid LSP typing for 12
Sub-millisecond structural queries; ~99% fewer tokens than file crawling
Single static binary — no Docker, no API keys, no telemetry
Installer auto-configures Claude Code and Cursor

Installation

Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.

Any MCP-compatible agent

Most agents (Claude, Cursor, Windsurf, VS Code, and more) read a standard mcpServers configuration.

Open your agent's MCP configuration file.
Merge the snippet below into it.
Restart the agent — the "Codebase Memory MCP" tools will be available.

{
  "mcpServers": {
    "codebase-memory-mcp": {
      "command": "codebase-memory-mcp",
      "args": []
    }
  }
}

Claude Code

Add the server with a single CLI command.

claude mcp add codebase-memory-mcp -- codebase-memory-mcp

Claude Desktop

Add the server to claude_desktop_config.json (Settings → Developer → Edit Config).

{
  "mcpServers": {
    "codebase-memory-mcp": {
      "command": "codebase-memory-mcp",
      "args": []
    }
  }
}

One-line install on macOS/Linux; add -s -- --ui for the graph visualiser.

{
  "command": "bash",
  "args": [
    "-c",
    "curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash"
  ]
}

Description

Ask an agent "who calls this function" and it will grep, open twenty files and burn six figures of tokens arriving at an answer a compiler would know instantly. Codebase Memory MCP replaces that with a graph: it indexes the repository once into nodes (functions, classes, routes) and edges (CALLS, IMPORTS, HTTP_CALLS) and answers structural queries in under a millisecond.

The indexing is two-layer. Tree-sitter provides fast syntactic parsing across 158 vendored grammars, pulling out definitions, imports and call sites. On top of that a hybrid LSP layer does type-aware refinement for a dozen languages — Python, TypeScript/JavaScript, PHP, C#, Go, C, C++, Java, Kotlin, Rust and Perl — resolving types across files without actually running a language server. Results are held in SQLite with LZ4 compression and written out in a single dump. A daemon coordinates sessions across multiple agents, runs background watchers and serves an optional graph visualisation on localhost.

The project's own measurement of the payoff: five graph queries consumed roughly 3,400 tokens where file-by-file exploration of the same question took about 412,000. It reports full-indexing an average repository in milliseconds and the Linux kernel — 28M lines across 75,000 files — in about three minutes.

It is a single static binary with no runtime dependencies, no Docker, no API keys and no telemetry; everything is local and code never leaves the machine. macOS, Linux and Windows builds are published, and the installer auto-detects Claude Code and Cursor and wires them up. MIT licensed.