codescope

A codebase knowledge graph for your AI agent. codescope indexes your repo into a local SQLite symbol graph and serves it over MCP — so Claude Code, Cursor, and Codex can answer “where is X, who calls it, what’s in this file” in a few tokens and one tool call, instead of grep-ing and reading whole files.

Coding agents burn tokens and tool calls re-discovering your code on every session. codescope does the discovery once, keeps it fresh as you type, and hands the agent the answer — not a pile of file contents.

  codescope

  ✓ indexed 262 files · 1,956 symbols · 23,881 refs  in 670ms
  ↻ watching for changes

  > search loadConfig
  function export loadConfig — src/config.ts:12  ·  function loadConfig(): Config

  > callers handleRequest
  src/server.ts:88   route
  src/worker.ts:41   processJob

  > affected src/store.ts
  3 test file(s) affected: test/store.test.ts, test/mcp.test.ts, test/indexer.test.ts

100% local

tree-sitter + SQLite. No API key, no network, no telemetry. Your code never leaves the machine.

Watch-first

A file watcher re-indexes each file on save in ~0.5 ms, so the agent always queries current code — never a stale snapshot.

21 languages

TypeScript, JS, TSX, Python, Go, Rust, Java, Ruby, C, C++, C#, PHP, Scala, Solidity, Zig, Kotlin, Objective-C, Lua, Bash, OCaml, ReScript.

Faster & leaner than the incumbent

Measured vs codegraph: 3.5–7.6× faster indexing, 3–5× smaller index, fewer tokens per answer, and higher caller accuracy. See benchmarks.

Install in one command

npx @abdulmunimjemal/codescope install

This wires codescope into Claude Code and Cursor automatically. Restart your agent and it can query your codebase graph.

Next steps

Quick start Index a repo and wire it into your agent in two minutes.

MCP tools The seven graph tools your agent gets, and when each is used.

Benchmarks Measured head-to-head vs codegraph — speed, size, tokens, and accuracy.

---

MIT   MCP server   github.com/abdulmunimjemal/codescope