code-atlas

Multi-language code intelligence MCP server — gives Claude Code (and any MCP client) a structured view of your codebase instead of raw text: symbol search, file outlines, AST pattern queries, cross-file references, call/type hierarchies, import graphs, change-impact analysis with affected-test detection, web-framework route maps (Express/NestJS/Fastify/FastAPI/Flask/Django), Mermaid diagrams, precise LSP-backed answers, local-embedding semantic search, and game-engine asset understanding (Godot, Unity, Unreal). 25 languages indexed.

Instead of grepping and reading whole files, the model asks questions like "map this repo", "outline this file", "who calls parseConfig?", "how does the request handler reach the DB layer?", "what changes whenever this file changes?" — and gets compact, token-efficient answers backed by a persistent tree-sitter index.

Every answer is budgeted: tools take a max_tokens cap, and a clamped response returns a cursor you can pass back to continue rather than re-asking with a bigger budget.

Status: released (npm @lusiem/code-atlas, MCP registry io.github.lusiem/code-atlas). 25 languages, 32 tools, LSP-exact answers with a structural floor, engine adapters, and an agent-workflow layer (repo_map, context_pack, verify_changes) built for token economy. Validated at 19k files / 6.3M LOC.

How it works

your repo ──scan (gitignore-aware)──> tree-sitter parse ──> SQLite index (symbols, imports,
                                                             occurrences, call/type edges, FTS5)
                                                                   │
Claude Code ◄──────────── MCP tools over stdio ────────────────────┘

• Persistent index at <repo>/.code-atlas/index.db (self-gitignored), incrementally refreshed by content hash.
• Live — a gitignore-aware file watcher reindexes saves within a debounce beat and re-resolves only the affected files (--no-watch or "watch": false to disable).
• Zero config — point it at a repo root and it works. Optional code-atlas.json for include/exclude tweaks.
• Everything local — your code never leaves your machine.

Install & use with Claude Code

claude mcp add code-atlas -- npx -y @lusiem/code-atlas serve

Or from a clone:

git clone https://github.com/lusiem/code-atlas && cd code-atlas
npm install && npm run build
claude mcp add code-atlas -- node /path/to/code-atlas/dist/index.js serve

The server indexes the current working directory; pass --root <path> to override.

CLI

node dist/index.js index [--root <path>]              # one-shot index build (debugging / warm-up)
node dist/index.js serve [--root <path>] [--no-watch] [--no-lsp] [--no-embeddings] [--no-download]
                                                      # MCP server on stdio

Tools

Full reference with example outputs: docs/tools.md.

| Tool | What it answers | |---|---| | repo_map | What is this repo and where do I start? Annotated directory tree with import fan-in/fan-out (layering at a glance), entry points, importance-ranked key symbols with the reason each ranks, and the third-party dependency inventory — one budgeted call instead of a grep sweep. path_prefix/depth to zoom. | | project_overview | The cheap version: languages, sizes, layout, index freshness. | | search_symbols | Where is X defined? FTS + fuzzy over names and doc comments, filterable by kind/language/path. | | search_content | Where does this text appear? String literals, error messages, env vars, config keys, TODOs — and the only tool that sees non-code files (md/json/yaml/toml/sql/Dockerfile/CI). The index proposes candidates, the disk verifies them, so a hit is never stale. | | semantic_search | "Where is retry backoff implemented?" — natural-language search, hybrid keyword+embedding ranking, fully local. | | get_file_outline | What's in this file? Hierarchical signatures without reading source — over budget it sheds nesting depth rather than tail lines, so every top-level symbol survives. Pass a directory for a package's public API surface in one call. | | get_symbol_info | Everything about one symbol (by id, position, or name) incl. docs and source. | | context_pack | One-call, token-budgeted briefing on a symbol: source, outline, callers/callees, types, route, related tests — instead of six lookups. Sections that don't fit the budget are named. | | batch_symbols | Up to 50 #ids resolved to compact one-liners in a single call. | | ast_query | Raw tree-sitter S-expression queries — structural search regex can't do. | | find_references | Who uses this symbol? Exact (LSP) when available, else resolved usages first, name-matches as candidates. | | go_to_definition | Definition of the identifier at a position. LSP-exact with index fallback. | | call_hierarchy | Who calls this / what does it call, as a tree. [lsp 1.00] edges when a language server is running. | | type_hierarchy | Supertypes and subtypes over extends/implements edges. | | get_dependencies | File import graph, both directions (imports / imported-by). | | trace_path | Shortest call chain between two symbols. | | change_impact | Blast radius of a change: transitive callers + import reachability, affected test files first. Target a symbol, a file list, or nothing — no args analyzes the uncommitted git diff (hunk-level: only symbols you actually touched seed the traversal). Affected route handlers are tagged [ROUTE GET /users/:id]. | | verify_changes | Post-edit structural check against git HEAD: imports that stopped resolving, removed exports other files still reference, signature changes with live callers. change_impact predicts; this confirms. | | tests_for_symbol | Which tests exercise this symbol? Reverse graph walk to test files, naming the test case; import-chain-only hits reported as a weaker signal. | | find_similar_code | "Does a helper for this already exist?" — near-duplicate search over the local embedding vectors, with a text-similarity fallback while coverage builds. | | find_dead_code | Zero-reference symbols after entry-point/route/lifecycle exclusions, confidence-hedged (possibly dead when a same-name usage exists anywhere); unused exports listed separately. | | hotspots | Churn × size risk ranking, answered from the cached commit history. | | change_coupling | Which files historically change together with this one — the coupling a structural index cannot see (config, fixtures, docs, string-keyed dispatch). Directional confidence, and files with no import edge are flagged. | | symbol_history | Why does this look like this, and who do I ask? git log -L over the symbol's line span: sha, date, author, subject. | | list_routes | Web-framework routes across the workspace — Express, Fastify, NestJS, FastAPI, Flask, Django, plus file-based routing for Next.js, SvelteKit, Nuxt, and Remix — each linked to its handler symbol. | | find_route | "Which code serves GET /api/users/7?" — matches a concrete URL against indexed route patterns (:id, {id}, <int:pk> are wildcards) and returns the handler. | | generate_diagram | Mermaid diagrams of the above graphs: import graph (file or directory level), call graph around a symbol, type hierarchy, call path — paste straight into GitHub markdown or docs. | | get_scene_structure | Godot scene node tree with attached scripts, instanced sub-scenes, and signal connections (handlers resolved to symbols). | | find_asset_references | Which scenes/prefabs/Blueprints use this script, asset, or handler? Reverse lookup across Godot res:// paths, Unity GUIDs (via .meta), and Unreal modules + /Game/ Blueprint paths. | | search_reflection | All UPROPERTY(Replicated), [SerializeField], @export vars, signals — engine reflection markers across the workspace. | | index_status / reindex | Index health and manual refresh. |

Cross-file answers are LSP-first with a structural floor: when a language server is available (found on PATH or auto-acquired into a per-user cache), references/definitions/hover/call hierarchies are exact and tagged lsp; everywhere else, heuristic import/name resolution answers with a confidence score per edge. Servers: typescript-language-server + pyright (npm), gopls (go install), rust-analyzer + clangd (pinned, SHA-256-verified release binaries), Eclipse JDT LS for Java (needs a Java 21+ runtime — detected via PATH/JAVA_HOME), kotlin-language-server (needs any JRE), and csharp-ls for C# (needs the .NET SDK; installed as a dotnet tool — deliberately not Microsoft's Roslyn LSP, whose license is VS-only). GDScript attaches to the Godot editor's built-in language server (TCP 127.0.0.1:6005, override with CODE_ATLAS_GODOT_LSP_PORT) whenever the editor has the project open — close the editor and answers fall back to structural, reopen it and the next query reattaches. Missing runtimes degrade that language to structural, reported honestly in index_status. A first-time acquisition (download + boot) never blocks a query: tools wait up to 8 s, then answer structurally while the server finishes starting. Disable with --no-lsp; disable auto-download with --no-download.

Semantic search embeds every function/class (signature + doc + body) with a code-tuned local model — jinaai/jina-embeddings-v2-base-code, quantized ONNX — and fuses cosine similarity with BM25 keyword rank. Everything stays on your machine. The ONNX runtime (~220 MB) and model (~150 MB) are not part of this package: they download to the per-user cache the first time you call semantic_search, never at install, and structural tools never wait on them. Until coverage completes, results are keyword-weighted and say so. Embedding a 60k-symbol repo takes ~15 minutes of background time, once; after that only edited symbols re-embed. "embeddings": {"model": "fast"} swaps in a 4× faster general-purpose model; --no-embeddings turns the layer off.

Languages

Indexing today (25): TypeScript, TSX, JavaScript, Python, C, C++, Rust, Go, Java, Kotlin, C#, GDScript, PHP, Ruby, Lua, Solidity, Zig, Nix, Swift, Scala, Dart, Terraform/HCL, Pascal/Delphi, Vue, Svelte.

Vue and Svelte single-file components index the <script> blocks (all of them, <script setup> included) at their true file line numbers; component imports resolve, so .vue/.svelte dependency graphs work in get_dependencies and change_impact. Terraform maps blocks to symbols (resource "aws_s3_bucket" "logs" is searchable by either label) and resolves local module sources as imports. Languages beyond the LSP-covered core answer structurally with confidence-scored edges — same contract as everywhere else.

Test-file awareness: every indexed file is classified by per-language path conventions (*.test.ts, test_*.py, *_test.go, src/test/, *_spec.rb, foundry .t.sol, …) — this powers change_impact's affected-test reporting. Inline test blocks (Rust #[cfg(test)], Zig test) are not path-classifiable and stay invisible by design.

Game engines: engine assets index alongside code — Godot .tscn/.tres scenes (node trees, script attachments, signal connections, autoloads; res:// resolved per project.godot, monorepos of many projects included), Unity .unity/.prefab/.asset + .meta GUID maps (MonoBehaviour → C# class links, serialized references), and Unreal .uproject/.uplugin/Build.cs module graphs plus indexed reflection macros: UCLASS/USTRUCT/UFUNCTION/UPROPERTY specifiers attach to the symbol they annotate (searchable, shown in get_symbol_info), and dllexport macros (class MYGAME_API AMyActor) no longer break C++ class extraction.

Unreal Blueprints (.uasset/.umap): binary editor packages are parsed natively (package summary, name/import/export tables, asset-registry tags, and bounded tagged-property reads — no Unreal installation needed). Blueprints surface their functions, event dispatchers, and member variables as searchable symbols (search_symbols kind:event_dispatcher, lang:uasset), a get_file_outline per asset, graph-node comments in full-text search, and asset//Game/-path + parent-class + called-function references in find_asset_references (including "who handles ReceiveBeginPlay" via handles_event). Blueprint→C++ links (parent classes via Unreal's A/U/F naming, called UFUNCTIONs) materialize as confidence-scored provenance: engine edges, lighting up type_hierarchy, call_hierarchy, and change_impact. Uncooked (editor) assets are the target — validated against a real UE 5.6 project (1,763/1,764 assets parsing, save vintages from UE 4.25 through 5.6); cooked/marketplace packages are detected and skipped with a recorded reason. Binary assets are capped by maxAssetFileBytes (default 64 MB) and change-gated by size+mtime so warm sweeps never re-read them. Godot binary .scn remains out of scope.

Roadmap

1. ~~Structural core: scanner, SQLite+FTS5 index, TS/JS/Python extractors, first 6 tools~~ ✅
2. ~~All 11 language extractors, cross-file import resolution, call graph (find_references, call_hierarchy, trace_path)~~ ✅
3. ~~File watcher + incremental reindexing (scoped re-resolution, schema migrations)~~ ✅
4. ~~LSP layer (auto-acquired ts-ls/pyright/gopls; checksum-pinned rust-analyzer/clangd binaries; JDT LS/kotlin-language-server/csharp-ls with JRE/.NET detection; precise references/definitions/hover/call hierarchy with graceful fallback)~~ ✅
5. ~~Local-embedding semantic search (semantic_search, hybrid BM25+vector reciprocal-rank fusion, lazy model download, incremental re-embedding)~~ ✅
6. ~~Game-engine adapters: GDScript grammar (vendored wasm build), Godot scenes/autoloads, Unity prefabs/GUIDs, Unreal module graph + reflection search, Godot editor LSP attach (TCP 6005, nightly-tested against a real editor)~~ ✅
7. ~~Docs, benchmarks in CI (cold index 157k LOC ≈ 5 s, warm queries p95 < 10 ms), cross-platform CI, npm pack smoke~~ ✅ — npm publish + MCP registry submission pending
8. ~~Competitive push: change_impact (blast radius + affected tests, git-diff mode), web-framework route indexing (list_routes/find_route), 13 more languages incl. Vue/Svelte SFCs~~ ✅
9. ~~Agent-workflow release (0.4.0): context_pack (token-budgeted one-call briefings), verify_changes (post-edit check vs git HEAD), tests_for_symbol, find_similar_code, find_dead_code + hotspots, batch_symbols, max_tokens on every list-shaped tool, file-based routing (Next.js/SvelteKit/Nuxt/Remix), PHP PSR-4 / C# namespace / Swift SPM import resolution, opt-in LSP edge promotion, sqlite-vec vec0 KNN~~ ✅
10. ~~Orientation release (0.5.0): repo_map (annotated architecture map, importance ranking, entry points, dependency inventory), resumable cursors + a recalibrated token estimator, depth-collapsing outlines and directory outlines, change_coupling + symbol_history on a persisted commit cache, streaming resolution for million-symbol repos~~ ✅
11. ~~Token-honesty and text release (0.6.0): an estimator that no longer under-counts high-entropy output (base64 was charged a quarter of its true cost, so max_tokens could be exceeded), provenance stated once per answer instead of on every row, search_content over a contentless FTS index with on-disk verification, non-code files (md/json/yaml/Dockerfile/CI) indexed for the first time~~ ✅

Performance (vuejs/core, 157k LOC, 536 files incl. .vue SFCs): cold index 3.8 s, warm tool calls p95 ≤ 6 ms through a full MCP round-trip. scripts/bench.mjs runs in CI.

Scale (Unreal Engine Engine/Source/Runtime, 19,097 files, 6.3M LOC, 622k symbols, 502k edges): cold index 413 s, warm structural queries p95 ≤ 18 ms, and repo_map returns the whole architecture in p95 126 ms within a 2,500-token budget. Text search over all 19k files costs p50 69 ms, and the content index is 45.6 MB — 4% of the database, because it stores no file text: it proposes candidates and the disk verifies them.

Development

npm install        # also copies grammar wasm files into grammars/
npm test           # vitest: extractor golden tests, store, scanner, MCP end-to-end
npm run lint
npm run dev serve  # run from source via tsx

Position convention: lines are 1-based, columns 0-based, paths root-relative with forward slashes. See CONTRIBUTING.md for how to add languages, servers, and grammars.

License

MIT