Gortex brings cross-repo code intelligence to MCP with 94% token reduction
Andrew Kumanyaev built Gortex, a Go-based tool that constructs an in-memory knowledge graph from multi-repo codebases and exposes it via 47 MCP tools, reporting a 94% reduction in token usage during complex investigations.
Score breakdown
Andrew Kumanyaev built Gortex after existing code indexers and graph-based retrievers failed to support cross-repository workflows. The tool constructs an in-memory knowledge graph capturing files, symbols, imports, call chains, and type relationships across multiple repos, then exposes this graph via 47 MCP tools that AI agents can query directly. A single `smart_context` call replaces the need to read 5–10 files to trace a call chain. Kumanyaev reports a 94% reduction in token usage — roughly 3,051,768 tokens saved across 618 calls — and cost avoidance ranging from ~$3 (claude-haiku-4.5) to ~$45 (claude-opus-4) based on built-in telemetry.
Andrew Kumanyaev describes building Gortex out of frustration with token-heavy AI coding sessions across multi-repository codebases. Existing tools required per-repo installation and re-indexing each time he switched between codebases, creating a fragmented workflow with constant context management overhead. Gortex addresses this with a workspace config that points at multiple repositories simultaneously, resolving symbols across all of them — the capability he found missing in every alternative he evaluated.
The knowledge graph Gortex builds goes beyond raw file contents, capturing files, symbols, imports, call chains, type relationships, and cross-repo dependencies.
The knowledge graph Gortex builds goes beyond raw file contents, capturing files, symbols, imports, call chains, type relationships, and cross-repo dependencies. It persists the graph to disk between sessions and restores it incrementally on startup. Key features include confidence-tiered call graph edges (flagging whether each edge was resolved by LSP, inferred from AST analysis, or matched via text patterns), community-scoped `SKILL.md` files auto-generated via Louvain clustering, and hybrid BM25 + vector semantic search powered by a bundled pure-Go ONNX runtime — no Python or external model server required. Indexing VS Code's ~10,700-file codebase takes about one minute on Apple Silicon; the Linux kernel (70K files, 1.69M nodes) takes around three minutes.
Kumanyaev reports that across 618 calls, Gortex returned 228,102 tokens while saving 3,051,768 — a 14.4x efficiency ratio. The tool exposes 47 MCP tools to AI agents, with `smart_context` as the primary interface for graph-derived answers. Gortex is written in Go, source-available under the PolyForm Small Business license, and free for individuals, open source projects, small businesses under 50 employees or $500K revenue, education, and government.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Apr 19, 2026 · 22:16 UTC. How this works →