Gortex (GitHub Repo)
GitHub
Gortex is a GitHub repo tool that maps code with tree-sitter and feeds agents only the bits they need. It says that cuts token use by up to 50x.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Gortex is trying to make AI coding assistants less chatty and more surgical. The GitHub repo says it parses 257 languages and grammars, builds a persistent knowledge graph of symbols and call chains, and lets agents query that graph instead of hauling entire files into context. The pitch is simple: ask for the bit that matters, not the surrounding noise.
That idea shows up everywhere in the project. Gortex ships 175 configurable MCP tools, zero dependencies, and support for 19 coding agents out of the box, from Claude Code and Cursor to VS Code / Copilot, Aider, Gemini CLI, and others listed in the docs. One install is supposed to configure whatever it detects on your machine. There’s also a Web UI for visualizing the graph.
The repo leans hard on speed and scale. It says a precomputed depth-3 reach index makes blast-radius queries cheap, and that the system can run across multiple repositories with evidence-gated resolution, contract matching, and impact analysis. There’s a claims stack here that covers more than code search: speculative execution, live editor overlays, semantic search, PR review, and a long-running daemon that serves every IDE window.
The benchmark section is the part that will get attention. On an Apple Silicon laptop, the project reports indexing torvalds/linux in about 3 minutes, microsoft/vscode in about a minute, and its own repo in 3.4 seconds, with parsing taking 65% to 80% of wall time. Gortex also publishes token-savings figures, including a dashboard that puts avoided cost at $168.69 for claude-opus-4 across 1,878 calls. That’s the kind of number AI tooling teams will absolutely want to scrutinize, and the kind of number this project clearly wants front and center.
My take — AI-written commentary, not fact-checked reporting
This is the right direction for coding agents: fewer files, more evidence, less theatrical page-flipping. The bigger story is that toolmakers are finally admitting context bloat is the real tax, not just model quality. Open graph-native plumbing beats another shiny wrapper over file reads, and the rest is mostly packaging.
Read more about this at: GitHub