mcp server

repotracer

A repo investigator for Claude Code and Codex. Offload codebase investigation and get up to 2.7× as much work from the same limits.

repotracer/repotracercommunitylistedunreviewednot running here yet

Use tool

Set up

Runs on your computer. Pick your agent and operating system.

This project does not publish an install method we can turn into a config. Follow the README instead.

README

synced Oct 15, 2026

Demo

npx repotracer@latest setup

What changes

Without RepoTracer, every search and file read stays in the same conversation and counts against the same limits the coding agent later uses to implement the change.

With RepoTracer, RepoTracer runs the investigation in a separate conversation and reports back to the coding agent.

We built RepoTracer to make the limits last longer. Then the code got better.

On the current complete-task benchmarks, the same limits covered 1.38× to 2.68× as much work. When the outputs of the two runs differed in quality, RepoTracer produced the better result.

Benchmarks

Task	Direct	With RepoTracer	Same limits	Implementation time
Production bug fix	Bug fixed	Bug fixed	2.68×	24.54% faster
SWE-bench Astropy 13453	Passing fix	Passing fix	2.00×	9.60% faster
Multi-language release	Broke 4 existing tests	Kept all existing tests passing	1.38×	16.83% slower

A complete run that costs 62.68% less lets the same fixed limit cover 2.68× as many runs.

How we measure

without RepoTracer = coding-agent usage
with RepoTracer    = coding-agent usage + investigator usage

We only call it a saving if the whole run costs less and still passes the task's quality check.

Raw runs, grading, and checksums are public.

Full benchmark writeup →
Raw benchmark artifacts →
Why complete-task measurement matters →

What the investigator can do

Trace behavior across files and related repos
Follow definitions, references, callers, and execution paths
Run checks or experiments to test a theory
Continue a previous investigation instead of starting over
Report what it could not resolve and what it already checked
Run independent investigations in parallel

When RepoTracer doesn't run

"Rename this variable in src/config.ts"
  → Claude Code / Codex

"Trace why refresh tokens fail after rotation"
  → RepoTracer

The router made the right call on all 42 cases in the current routing test.

Install it once. Prompt normally.

Claude Code and Codex

Choose Claude Code, Codex, or both.

RepoTracer uses the native CLI you are already signed into. No second provider login or gateway account is required.

Integration details →

Bring your own investigator

Use any OpenAI-compatible endpoint, including Ollama or vLLM.

npx repotracer@latest setup \
  --base-url http://localhost:11434/v1 \
  --model deepseek-coder

You can. The basic idea is the same: let another agent investigate the repo.

With RepoTracer, Claude Code or Codex can call that investigator through one MCP tool: repo_scout. Follow-ups can continue the same investigation instead of starting over.

It says so. The report includes what it found, what it checked, and what remains unresolved. Claude Code or Codex can continue from that work instead of repeating the investigation from scratch.

Install

npx repotracer@latest setup

Choose Claude Code, Codex, or both, then accept the defaults or pick your investigator.

Run the same command later to reconfigure or uninstall.

Want the CLI in your shell too?

npm install -g repotracer
# or
cargo install --git https://github.com/repotracer/repotracer --locked repotracer

Architecture

RepoTracer is a Rust MCP server. Native investigations run through the Claude Code or Codex CLI you already use. For OpenAI-compatible models, RepoTracer provides the repo tools and drives the investigation itself.

Full architecture →

repotracer "where is authentication handled?"
repotracer scout "trace refresh rotation"
repotracer symbols "Config" --mode references
repotracer serve
repotracer doctor
repotracer status
repotracer settings
repotracer update
repotracer uninstall --yes

serve runs the MCP server over stdio. symbols performs a local syntax lookup without a model call.

Develop

git clone https://github.com/repotracer/repotracer
cd repotracer
cargo test --workspace
cargo run -p repotracer -- doctor
cargo run -p repotracer -- scout "where is config loaded?" --mock

CI runs without a GPU, local model runtime, or API key. Live native-CLI smoke tests are documented in CONTRIBUTING.md.

Resources

RepoTracer is an independent project inspired by FastContext. It is not affiliated with or endorsed by Microsoft. See NOTICE.

License

MIT. See LICENSE and NOTICE.