One binary. Zero dependencies. 29 languages. 32 MCP tools. Local code intelligence for AI coding agents.
Documentation · npm · Issues
Any MCP-compatible AI agent, IDE, or CLI tool connects out of the box. cortex install auto-detects your setup and writes the config.
All 25 supported platforms
| Platform | Install command |
|---|---|
| Claude Code (Linux/Mac) | cortex install |
| Claude Code (Windows) | cortex install --platform claude-code |
| Codex | cortex install --platform codex |
| OpenCode | cortex install --platform opencode |
| GitHub Copilot CLI | cortex install --platform copilot |
| VS Code Copilot Chat | cortex vscode install |
| Aider | cortex install --platform aider |
| OpenClaw | cortex install --platform openclaw |
| Factory Droid | cortex install --platform droid |
| Trae | cortex install --platform trae |
| Trae CN | cortex install --platform trae-cn |
| Gemini CLI | cortex install --platform gemini |
| Hermes | cortex install --platform hermes |
| Kimi Code | cortex install --platform kimi |
| Kiro IDE/CLI | cortex kiro install |
| Pi coding agent | cortex install --platform pi |
| Cursor | cortex cursor install |
| Google Antigravity | cortex antigravity install |
| Windsurf | cortex install --platform windsurf |
| Zed | cortex install --platform zed |
| JetBrains | cortex install --platform jetbrains |
| Cline/Roo | cortex install --platform cline |
| Continue.dev | cortex install --platform continue |
| Supermaven | cortex install --platform supermaven |
| Tabnine | cortex install --platform tabnine |
An agent asking "what calls processOrder" gets a 200-token graph result instead of burning 20,000 tokens reading files. That's 100x fewer tokens on a single structural question.
Cortex indexes your repository into a SQLite call graph and exposes it over the Model Context Protocol. 127 files indexed in 535ms. Incremental re-index in 13ms. Handles 100K to 1M+ line repositories.
v1.1 adds multi-source confidence-aware intelligence. Cortex now ingests SCIP indexes for precise symbol resolution, runs framework-specific adapters (FastAPI, Express, NestJS, Spring, Django, React) to detect dependency injection and routing wiring, tags every edge with a confidence tier, and falls back honestly when it can't give a confident answer. The ask and get_task_context tools are completely overhauled with evidence-fusion ranking and structured fallback suggestions.
npx @1337xcode/cortex@latest installOther methods
Shell script
curl -fsSL https://raw.githubusercontent.com/1337Xcode/cortex/main/install.sh | shBuild from source
git clone https://github.com/1337Xcode/cortex.git &&cd cortex
cargo build --release
cp target/release/cortex ~/.local/bin/cortex index # parse your repo, build the graph
cortex install # detect Claude Code / Cursor, write MCP config
cortex serve # start the MCP serverYour agent queries the graph instead of reading raw files.
cortex impact UserService.getUser # blast radius: what breaks if I change this?
cortex explain DatabasePool.acquire # offline function explanation, zero LLM calls
cortex security report # taint flows, OWASP patterns, dependency vulns
cortex diff main feature-branch # call graph diff between branches
cortex viz --export graph.html # interactive 3D graph in a standalone HTML file
cortex ci --fail-on-taint # CI quality gate with exit codes
cortex hotspots --months 6 # high-churn risky code ranked by risk score
cortex coverage --lcov coverage.lcov # populate coverage field, rank untested functions
cortex modules # Leiden community detection module boundaries
cortex federate add ../auth-service # query across repos with unified graph
cortex benchmark # run correctness suite, report pass rate + token savings
cortex status --savings # token savings dashboard with honest net-negative reporting
cortex index --repair # rebuild index from scratch, preserve observations
cortex semantic enable# build embedding index for semantic search- 29 languages parsed via tree-sitter
- 32 MCP tools exposed over stdio, or 5 in smart mode (the
askmeta-tool routes internally) - SCIP index ingestion for precise symbol resolution — SCIP edges win over tree-sitter on conflicts
- 6 framework adapters (FastAPI, Express, NestJS, Spring, Django, React) detect DI wiring, middleware chains, and routing that tree-sitter cannot see
- Confidence-tagged edges — every edge carries
edge_sourceandconfidence(1.0 SCIP / 0.8 framework / 0.5 AST / 0.3 name-match); queries default to confidence ≥ 0.7 - Evidence-fusion
get_task_context— multi-signal ranking (BM25 + SCIP distance + git recency + edge confidence + file size penalty) with per-file reasons and guaranteed non-empty results - Honest fallback —
askreturns structured grep suggestions when confidence is below MEDIUM; never returns confidently wrong data from an unhealthy index - Index health gate — all tools return a structured error with fallback suggestions when the index is empty or corrupt
- Token savings dashboard —
cortex status --savingsshows cumulative net saved, average per query, and net-negative queries (reported honestly) get_repo_brief— zero-parameter cold-start summary under 400 tokens: languages, frameworks, entry points, hotspots, security patterns, test shape- Tool surface management — 10 default tools, 7 experimental (opt-in), 5 smart-tools mode;
semantic_searchonly listed when embeddings are built - Incremental embeddings — only re-embeds functions whose content hash changed;
cortex semantic enablebuilds the initial index - Correctness benchmarks —
cortex benchmarkvalidates tool accuracy against ground-truth; CI gate fails the build if pass rate drops below 70% - Sub-second incremental re-indexing via native OS file watcher (inotify, FSEvents, ReadDirectoryChangesW)
- Configurable model pricing via
~/.cortex/pricing.tomlwith longest-prefix matching - Taint flow analysis, OWASP Top 10 detection, SBOM generation all running locally, no cloud
- Cross-session memory that marks observations stale when linked code changes
- Multi-repo federation for querying across repositories with a unified graph
- CI quality gates with configurable thresholds and exit codes
- 3D interactive graph visualization exportable to standalone HTML
- Portable JSON bundle (cortex.json) for team sharing via git
- Works offline. No cloud, no API keys, no Docker, no language runtimes.
All 29 supported languages
| Language | Language | Language |
|---|---|---|
| Python | TypeScript | JavaScript |
| Rust | Go | Java |
| C | C++ | C# |
| Ruby | PHP | Swift |
| Kotlin | Scala | Lua |
| Zig | Haskell | Elixir |
| Dart | R | Julia |
| OCaml | Bash | HCL/Terraform |
| Perl | Objective-C | SQL |
| YAML | TOML |
graph TD
A[Repository Files] --> B[File Watcher<br/>notify crate]
B --> C[Indexer Pipeline<br/>Rayon + tree-sitter]
C --> D[SQLite Graph Store<br/>WAL + FTS5 + sqlite-vec]
D --> E[MCP Server<br/>Tokio + JSON-RPC 2.0]
E --> F[AI Coding Agent]
C --> G1[SCIP Ingester<br/>Protobuf → HIGH edges]
G1 --> D
C --> G2[Framework Adapters<br/>FastAPI · Express · NestJS<br/>Spring · Django · React]
G2 --> D
C --> G3[Security Pass<br/>Taint + OWASP + SBOM]
G3 --> D
D --> H[Bundle Exporter<br/>cortex.json]
H --> I[Git Repository<br/>Team Sharing]
D --> J[Memory Layer<br/>Staleness-aware observations]
D --> K[Evidence Fusion<br/>BM25 + SCIP + git + embeddings]
K --> E
Every edge in the graph carries a source and confidence value. Queries default to confidence ≥ 0.7 (MEDIUM), filtering out heuristic name-match edges.
| Source | Confidence | How produced |
|---|---|---|
scip | 1.0 HIGH | SCIP index (precise symbol resolution) |
framework_adapter | 0.8 MEDIUM | FastAPI/Express/NestJS/Spring/Django/React pattern matching |
ast_direct | 0.5 LOW | tree-sitter AST extraction |
name_match | 0.3 VERY_LOW | heuristic name-based resolution |
Cortex ships with a self-referential benchmark suite (benchmark/cortex_self_benchmark.json, 24 cases) covering trace_callers, blast_radius, get_task_context, and ask. Run it against your index:
cortex benchmarkThe CI pipeline runs this on every release and fails the build if pass rate drops below 70%.
| Feature | Cortex | codebase-memory-mcp | Srclight | Serena | mcp-codebase-index |
|---|---|---|---|---|---|
| Architecture | Rust binary, SQLite | C binary, SQLite | Python, SQLite | Python, LSP | Python, SQLite |
| Languages | 29 (tree-sitter) | 66 (tree-sitter) | 10 (tree-sitter) | 40+ (LSP) | 4 |
| MCP tools | 32 (or 5 smart mode) | 14 | 42 | ~20 | 17 |
| Smart tool routing | Yes (ask meta-tool) | No | No | No | No |
| SCIP precise resolution | Yes | No | No | Yes (LSP) | No |
| Framework adapters | 6 (DI, routing, middleware) | No | No | No | No |
| Confidence-tagged edges | Yes (4 tiers) | No | No | No | No |
| Honest fallback | Yes (FallbackSuggestion) | No | No | No | No |
| Index health gate | Yes | No | No | No | No |
| Token savings dashboard | Yes (honest math) | No | No | No | No |
| Cold-start repo brief | Yes (get_repo_brief) | No | No | No | No |
| Correctness benchmarks | Yes (CI gate, 70% threshold) | No | No | No | No |
| Call graph | Function-level | Full chains | Callers/callees | LSP-precise | Cross-file |
| Hybrid search | FTS5 + sqlite-vec | Graph only | RRF fusion | Keyword | Structural |
| Token reduction | 100x on structural queries | 121x avg | Not measured | Not measured | 87% |
| Incremental update | 13ms (no changes) | Milliseconds | Git hooks | LSP live | Hash-based |
| Security (taint/OWASP/SBOM) | Yes | No | No | No | No |
| Cross-session memory | Staleness-aware | No | No | No | No |
| Multi-repo federation | Yes | No | SQLite ATTACH | Partial | No |
| Community detection | Leiden algorithm | No | Leiden | No | No |
| Coverage gap analysis | Yes (LCOV cross-ref) | No | No | No | No |
| Single binary, zero deps | Yes | Yes | No (Python) | No (Python+LSP) | No (Python) |
| Auto IDE config | 25 agents | No | No | No | No |
| Air-gap compatible | Yes | Yes | Yes | Yes | Yes |
| License | MIT | MIT | MIT | MIT | MIT |
Full docs site: 1337xcode.github.io/cortex
Documentation · npm · Issues