Gortex

Gortex is an open-source code-intelligence engine built by Andrey Kumanyaev (zzet) and released under Apache 2.0. It indexes a repository into an in-memory knowledge graph and exposes that graph to AI coding agents through a 21-tool MCP surface, a versioned HTTP API, a CLI, and a web UI — all from a single static binary with zero external dependencies. The project is actively developed on GitHub (zzet/gortex) and is currently at v0.63.8.

What It Is

Gortex sits between your codebase and your AI coding agent. Instead of letting the agent issue a series of file reads, greps, and globs to understand context, Gortex answers the same question in one graph query — returning callers, callees, imports, blast radius, and interface implementors in a single response. The project claims this approach delivers up to 50× fewer tokens per response compared to naive file reads, with reproducible benchmarks published in the repository. It supports 257 languages through a three-tier extraction pipeline (bespoke tree-sitter for deep resolution, regex for niche/legacy languages, and forest-backed signature-only for the long tail), and wires into 19 AI coding agents out of the box.

Architecture and Deployment Model

Gortex ships as a single static binary (~20 MB, CGO-on) for macOS, Linux, and Windows on both amd64 and arm64. Everything runs in-process and in memory — no database to run, no container to orchestrate. Key internal components include:

• Knowledge graph core — nodes for files, symbols, imports, calls, fields, SQL tables, K8s resources, Dockerfile images, and more; edges computed with Tarjan SCC, Leiden/Louvain communities, and a precomputed reach index for sub-millisecond blast-radius queries.
• Long-living daemon — a Unix-socket daemon holds one graph per tracked repo, with live fsnotify watching, per-session isolation, and launchd/systemd auto-start.
• LSP bridge — wired to 17 language servers (gopls, tsgo, pyright, rust-analyzer, clangd, jdtls, and others) for compiler-grade resolution.
• MCP server — stdio and Streamable HTTP (MCP 2026) transports; named agents get a compact 21-tool surface by default, with a 175-tool catalogue available on demand.
• GCX1 wire format — a published, round-trippable compact text format that the project reports delivers a median −27.4% token reduction versus JSON at the same fidelity.

Releases are Sigstore-signed with SLSA Level 3 build provenance and tracked by OpenSSF Scorecard.

Agent Integrations

gortex init auto-detects every supported coding agent on the machine and writes the correct configuration for each. The project lists 19 first-class adapters:

• Claude Code (slash commands, skills, hooks)
• Cursor, Windsurf, VS Code / Copilot, Continue.dev, Cline, Zed, Aider
• Kiro, OpenCode, Codex CLI, Gemini CLI, Antigravity, Kilo Code
• OpenClaw, Hermes, Oh My Pi, Pi, Kimi

Each adapter writes agent-specific config files (.mcp.json, CLAUDE.md, .cursor/rules, .windsurfrules, etc.) and community-routing blocks so the agent knows how to use Gortex's graph tools.

Key Capabilities

Beyond basic symbol lookup, Gortex exposes a broad set of analysis tools:

• Speculative execution — change(preview) and change(simulate) answer "what changes if I apply this edit?" without touching disk.
• Live editor overlays — push unsaved buffers as a shadow graph; all subsequent tool calls read through a per-request layered view.
• Dataflow and taint analysis — trace(flow) returns ranked dataflow paths; trace(taint) runs source→sink sweeps for security audits.
• Clone detection — MinHash + LSH banding emits similar_to edges at index time.
• Cross-repo API contracts — auto-detected across HTTP, gRPC, GraphQL, Kafka/NATS/Redis pub/sub, WebSocket, OpenAPI, and Temporal workflows; normalised to canonical IDs and matched provider↔consumer across repos.
• Graph-grounded PR review — a deterministic rulepack (NPE, thread-safety, N+1, logic-error) runs over a changeset and emits a BLOCK/REVIEW/APPROVE verdict with line-anchored comments.
• Semantic search — hybrid BM25 + vector with RRF fusion; default-on with a baked GloVe-50d table (3.8 MB embedded, CPU-only); opt-in MiniLM, Ollama, or OpenAI.
• Token savings dashboard — gortex savings reports tokens saved versus naive file reads per-call, per-session, and cumulatively.

Update: v0.63.8

The latest release is v0.63.8, published on 2026-08-20. Recent changelog entries on the homepage (since v0.58) include: indexer and resolver SQLite hot-path performance sweep, Merkle parsed-source reuse, lazy clone sketches, Terraform Lambda env var contract providers, Java Type::new and qualified method reference support, overloaded method ref binding, layered .gitignore handling, Visual Studio/MSBuild artifact exclusion, test-file exemption from coverage debt in PR review, the explore(localize) tool, a compact 21-tool MCP surface for named agents, instruction profiles (core/localization/full), the gortex guide verb and gortex://guide resource, and Kimi + Pi adapters bringing the total to 19 wired agents. The project has 1,487 GitHub stars and 135 forks as of the last recorded update.