Travsr

The Code Graph that lives next to git.

Problem: AI coding tools today depend on grep, which matches strings or vector RAG which uses approximate similarity to understand codebases. Neither method knows about a project’s call graph, types or imports. As a result when an agent asks about `charge()` the tool returns unrelated matches and can even miss the correct one if the function has been renamed.

What We built: Travsr is a code graph that lives locally in your repository. Travsr is exposed to AI agents through MCP or Model Context Protocol. Travsr gives agents precise reasoning about a codebase, rather than guesses based on text similarity.

How it works:

- tree-sitter parses each file once. It does not rebuild the full AST.

- A semantic pass takes the parsed data and resolves call targets, types and data flow. The results are stored in a SQLite graph at `.travsr/graph.db`.

- Agents query the graph through MCP. Nodes are ranked with PageRank. The results are returned pre-tokenized for the model.

- A git hook keeps the index fresh incrementally. It is fast enough to run on every commit.

Key features:

- 26 first‑party MCP tools, using stdio or SSE. They include index‑freshness and graph‑health checks.

- Structural support for 16 languages.

- Deep semantic support for TypeScript, JavaScript, Rust and Python. You can extend support by running `travsr lang install`.

- A VS Code extension shows a Cytoscape graph view, graph statistics and synonym discovery.

Stack: tree‑sitter, SQLite, MCP, TypeScript and Rust. Adjust the list to reflect what is actually true.

Status: Open source, under the Apache‑2.0 license.

Links: GitHub · www.travsr.com · docs.travsr.com