SourceMinder

SourceMinder is a command line tool that indexes your codebase by symbol so AI agents and language models can search it efficiently without shipping the whole thing to a cloud service. It parses source files locally using Tree-sitter, builds a symbol database, and lets you query that database in milliseconds. The goal is reducing the token overhead that comes with naive approaches like grepping and concatenating files.

The problem it addresses is familiar to anyone who has tried to point an LLM at a large codebase. You can dump files into context, but that burns tokens fast and often sends irrelevant code. SourceMinder builds an index once, then queries return just the symbols you need, typically in around six milliseconds across a quarter million symbol index. The claim is roughly fifteen percent fewer tokens compared to grep and cat workflows.

It runs entirely on your machine. There are no cloud calls, no API keys, and no data leaving your network. The index lives locally, and you query it with a tool called qi that supports wildcards, type filters, and options to expand definitions or exclude comments. You build the index with language specific commands, one per supported language.

Language support currently includes C, Go, Perl, PHP, Python, TypeScript, and Rust, with Rust and Perl marked as provisional. The underlying parser is Tree-sitter, which means adding more languages is possible but depends on the maintainers or community contributions.

Installation is a single curl command that pulls a prebuilt binary. It ships ready for x86_64 Linux, with support for macOS and BSD as well. The project is licensed under GPL v3, so you can inspect the source and contribute if you want to extend it.

The audience is developers who use AI coding assistants or build their own agent workflows and want tighter control over what context gets sent. If you have a large codebase and are tired of watching token counts balloon, SourceMinder offers a way to query for exactly what you need instead of grabbing everything and hoping the model filters it out.

It is free and open source. The current version is 0.3.1 beta, so it is still maturing, but the core indexing and querying workflow is functional and documented on GitHub.