Girder cuts agent file reads by 97.85%

Girder answers an agent's code question with 8,765 bytes where a full-file read costs 408,137, per its own ten-node test case. The count is bytes, not tokens. That's still a 97.85% cut in what an agent has to read before it can act. For anyone running agents against large repos, it means less of the context window spent on file contents the task doesn't need.

The tool parses a repo into functions, definitions, and call edges. It runs as one static Rust binary and answers queries over MCP. Beyond raw source lookups, it traces the call graph for impact analysis and picks the tests that cover a changed function. It can also apply edits addressed by graph location instead of by line number.

The free tier covers get_source, find_definition, search_code, ask_codebase, and review_changes. License checks run offline, and there's no telemetry.

Rust and Python parsing is mature. TypeScript and Go are still gated behind documented limits, so repos in either language won't see close to that byte cut yet.