Gortex

Gortex indexes repositories into a persistent code knowledge graph and serves it to coding agents over MCP, an HTTP API, a command-line interface and a web interface. Parsing covers 257 languages and grammars through tree-sitter in three tiers — bespoke grammars for deep resolution, regular expressions for niche languages, and signature-only extraction for the long tail — with compiler-grade resolution for Python, TypeScript, JavaScript, PHP, C#, Go, C, C++, Java, Kotlin, Swift, Zig, Rust, Ruby, Elixir, OCaml and Haskell. The graph holds functions, classes, call chains, HTTP routes and cross-service contracts, and each call edge records how it was resolved.

The purpose is token accounting: one graph query returns the symbols, callers, callees, imports and blast radius an agent would otherwise assemble from several file reads, which the project measures at up to 50x fewer tokens per response, with a further reduction from its own wire format compared with JSON. A precomputed depth-three reach index turns blast-radius questions into map lookups, and multiple repositories share one graph so references, contracts and call chains cross repository boundaries.

It ships as a single static binary for macOS, Linux and Windows with no external dependencies and no model download to start. A long-running daemon serves every client from an on-disk SQLite store, watching the filesystem for changes and restarting incrementally, and one setup pass configures every coding assistant detected on the machine. Semantic search is on by default with a small embedded word-vector model and full-text ranking fused adaptively, with larger embedding backends optional. Telemetry is off unless enabled.

Cross-repository contracts are detected and matched between providers and consumers across HTTP routes, gRPC services, GraphQL schemas, message topics on Kafka, RabbitMQ, NATS or Redis, WebSocket events, environment variables, OpenAPI specifications and Temporal workflows, normalised to canonical identifiers so orphan providers, orphan consumers and mismatches surface.

Repository	Files	Nodes	Edges	Index time	Peak heap
torvalds/linux	70,333	1,690,174	6,239,570	~3 min	5.07 GB
microsoft/vscode	10,762	204,501	808,902	~1 min	580 MB
the project's own repository	430	5,583	53,830	3.4 s	52 MB

Features

• 257 languages: three-tier extraction plus drop-in user grammars, and Jupyter and Databricks notebooks
• Cross-repository graph: many repositories in one graph, with evidence-gated resolution and per-session isolation
• Blast radius: a precomputed reach index answers what-breaks-if-I-change-this on every edit
• Contract matching: HTTP, gRPC, GraphQL, message topics, WebSocket, environment variables, OpenAPI and Temporal, matched provider to consumer
• MCP surface: a compact default tool set for named agents, with the rest of a 175-tool catalogue reachable through search
• Speculative execution: preview and simulation tools answer what a proposed edit would change without touching disk
• Editor overlays: unsaved buffers pushed as a shadow graph, with branching for parallel speculative sessions
• Graph-grounded review: per-pull-request blast radius, merge-order conflicts, a ranked queue and line-anchored findings with a verdict
• Semantic search: an embedded word-vector model with full-text ranking and adaptive fusion, with larger backends optional
• Daemon and interfaces: one process serves every client over stdio, HTTP and a graph web interface with filesystem watching
• Token savings reporting: per-call, per-session and cumulative token savings priced against a headline model
• Privacy defaults: telemetry off unless enabled, honouring the do-not-track environment variable