codescope

Local-first codebase knowledge-graph MCP server. Parses your repo into a symbol graph and serves it to AI coding agents so they stop wasting tokens re-scanning files. Watch-first: the graph stays fresh as you type.

Coding agents (Claude Code, Cursor, Codex, …) burn tokens and tool calls re-discovering your codebase — grep for a name, read the whole file, grep again for the callers, read those files too. codescope indexes the repo once into a local SQLite graph and answers "where is X, what calls it, what's in this file" in a handful of tokens and a single tool call — then keeps the graph current by re-indexing each file the instant you save it.

100% local. No API keys, no network, no telemetry.

Why not just grep?

grep finds text; codescope understands structure. It knows that run is a method on Service, that loadConfig is called from three places, and that a bare parse() call is a different thing from obj.parse(). It returns file:line + signatures, not raw matches — and it returns a bounded call neighbourhood (callers + callees, a few hops out) so an agent gets the relevant slice of the codebase for a change without opening a dozen files.

See BENCHMARKS.md: on a 2,500-file repo, codescope answers a navigation query in ~70–98% fewer tokens than reading the file, and refreshes a changed file in ~0.5 ms — roughly 3,000× cheaper than a full re-index.

How it compares to codegraph

codegraph (~35k★) is the mature incumbent and shares codescope's architecture. In a measured head-to-head (BENCHMARKS.md, both tools run on the same repos), codescope wins the efficiency axes:

• More accurate answers. Scored against the TypeScript compiler as ground truth (bench/accuracy.mjs), codescope's callers beat codegraph's F1 on every package tested (0.95 vs 0.66, 0.92 vs 0.70, 0.96 vs 0.91) — it never misses a true caller (recall 1.00) where codegraph misses 10–35%.
• 3.5–7.6× faster indexing (670 ms vs 2,335 ms on 262 files; 2.6 s vs 20 s on 3,500) — parsing is fanned across a worker-thread pool.
• 3–5× smaller index on disk (2.5 MB vs 8.2 MB; 22.8 MB vs 112.8 MB).
• Fewer tokens per answer — on both definition and callers queries, every repo tested.
• Feature parity: callers, callees, impact, context, affected (test-impact), and install (agent auto-wiring) — across 21 languages.

codegraph still leads on maturity & adoption (35k★, a real user base) and a few extra node kinds (constants, properties, routes). Pick codescope when accuracy, footprint, index speed, and token cost matter; pick codegraph for the most battle-tested option. Full numbers in BENCHMARKS.md.

Install

npx @abdulmunimjemal/codescope mcp .          # zero-install, or:
npm i -g @abdulmunimjemal/codescope           # then the `codescope` command is on your PATH

Requires Node ≥ 18. (The npm package is scoped because the bare name codescope collides with an existing package; the installed command is still codescope.)

Quick start

The one-liner — wire codescope into your agents automatically, from your repo:

npx @abdulmunimjemal/codescope install     # adds codescope to Claude Code + Cursor

That writes the MCP server config (non-destructively) so your agent launches codescope on the repo. Restart the agent and you're done. --agent claude|cursor targets one; --global writes to your home dir instead of the project.

Prefer to wire it by hand? The server command is:

codescope mcp /path/to/your/repo            # index, watch, and serve over stdio

Claude Code (.mcp.json or claude mcp add):

{
  "mcpServers": {
    "codescope": { "command": "npx", "args": ["-y", "@abdulmunimjemal/codescope", "mcp", "."] }
  }
}

Cursor / Codex / any MCP client: use the same command — npx -y @abdulmunimjemal/codescope mcp . over stdio.

You can also drive it straight from the terminal:

codescope index .                       # build the graph, print stats
codescope search useState               # fuzzy symbol search
codescope get GraphStore                # jump to a definition
codescope callers parseSource           # who calls this
codescope callees indexAll              # what it calls
codescope impact GraphStore             # blast radius before a change
codescope context "auth flow"           # ranked relevance map for a task
codescope affected src/store.ts         # which tests are affected by a change
codescope neighborhood handleRequest --depth 3
codescope watch .                       # keep the graph fresh, log updates

MCP tools

tool	what it answers
search_symbols(query, kind?, limit?)	fuzzy substring search over definitions — use instead of grep/glob
get_symbol(name, limit?)	jump to a definition by exact name (kind, file:line, signature)
find_callers(name, limit?)	who calls this function/method (distinct callers)
find_callees(name, limit?)	what this symbol calls — its outgoing dependencies
impact(name, depth?, limit?)	transitive callers (blast radius) before you change something
context(query, maxSymbols?)	a ranked relevance map for a task — matches + neighbours, the fastest way to orient
find_references(name, kind?, limit?)	all calls + imports of a name
file_outline(path)	every symbol in a file, in order — a compact alternative to reading it
neighborhood(name, depth?, limit?)	the call neighbourhood (callers + callees) around a symbol, as a subgraph
stats()	counts for the indexed graph

Tool descriptions are written for the agent — they nudge it to query the graph instead of scanning files.

How it works

1. Parse. Every supported file is parsed with tree-sitter (WASM grammars, no native build) into definitions (functions, methods, classes, interfaces, types, enums) and references (calls, imports).
2. Store. Symbols and references go into a local SQLite database with a trigram FTS5 index for fast substring search. References are stored by name and resolved to definitions lazily at query time — so changing one file never invalidates another's data.
3. Resolve. Calls are resolved kind-aware: a bare foo() resolves to a function named foo, while x.foo() resolves to a method named foo. This avoids the classic name-collision explosion (e.g. a project that happens to define a function called push). Ambiguous, library-ish names are left unresolved rather than blowing up the graph.
4. Watch. A file watcher re-indexes each file on save in sub-millisecond time. Because updates are per-file and content-hash gated, the graph is always current and a re-scan skips everything that hasn't changed.

The index lives in .codescope/graph.db (add .codescope/ to your .gitignore). codescope respects your repo's .gitignore when indexing.

Languages

21 languages: TypeScript, JavaScript, TSX/JSX, Python, Go, Rust, Java, Ruby, C, C++, C#, PHP, Scala, Solidity, Zig, Kotlin, Objective-C, Lua, Bash, OCaml, and ReScript. Definition extraction (functions, classes, methods, …) works for all; call and import edges are available for the languages whose grammars expose them.

Programmatic API

Everything is importable:

import { GraphStore, Indexer, watch, parseSource } from "codescope";

const store = new GraphStore("graph.db");      // or ":memory:"
const indexer = new Indexer(store, "/repo");
await indexer.indexAll();

store.searchSymbols("config");
store.neighborhood("handleRequest", { depth: 2 });

watch(indexer, { onChange: (file, action) => console.log(action, file) });

Limitations

• References resolve by name + call shape, not full type/scope analysis. It is a fast heuristic graph, not a compiler. Cross-file import resolution is not yet modelled.
• Rust impl methods are currently labelled function (impl blocks aren't tracked as containers).
• Symbol extraction targets top-level and class-member definitions; deeply nested local helpers are captured, anonymous expressions are not.

Roadmap

• Cross-file resolution — resolve an imported callee to its specific definition file (and, eventually, type-aware method resolution) to push precision toward 1.0.
• More languages — the language system is config-driven; new grammars are a table entry plus a test. Open a request or send a PR.
• Richer nodes — optionally index constants, properties, and routes.

Contributing

Contributions are very welcome — codescope is small, fully tested, and designed to be easy to extend. Adding a language is often a single config object plus a test. See CONTRIBUTING.md for setup, the project layout, and a step-by-step "add a language" guide, and please follow the Code of Conduct.

pnpm install && pnpm test && pnpm typecheck && pnpm build

Changelog: CHANGELOG.md · Security: SECURITY.md

License

MIT © Abdulmunim Jemal