Code Finder: fast, efficient code search for coding agents

Code Finder is an agentic search tool in the Sourcegraph MCP server, now available in Beta. You give it a task, it runs its own search loop across a repository, and it returns the matching file paths and line ranges with a short note on what each one does. Your coding agent reads a focused answer and moves on.

When a coding agent searches on its own, it spends turns opening files, scanning them, and backing out of dead ends, and every one of those turns eats into the context window it needs for the actual work. Code Finder takes the task, does the searching, and returns only the parts of the codebase that answer it.

What the benchmark shows

To measure the difference, we assembled a suite of file-finding tasks across repositories ranging from small to very large. We ran each task three ways: Code Finder, a coding agent calling our MCP tools, and a coding agent using its own local search tools.

Code Finder finished more than twice as fast as the agent using local search, and cost up to 40% less than the agent calling our MCP tools, at comparable result quality. The numbers below are relative to Code Finder, so lower is better. See the changelog post here.

The gains come from division of labor. A general coding agent carries a large context window and a broad instruction set into every search. Code Finder is built for one job, so it can search a repository more directly and return a smaller, denser answer for the agent to act on.

Try it

Code Finder is free while in Beta, and it already powers file discovery inside Deep Search, where it lets Deep Search spend less of its token budget locating files and more on reasoning about the code. To use it from your own agent, connect the Sourcegraph MCP Server. For a short walkthrough, watch the changelog video.