repotracer/repotracer: codex subscriptions have a fixed monthly budget

Search with Luna. Code with Sol. Never burn turns finding files.

Make your Codex subscription last up to 2.7x longer. A cheaper model searches your repo via MCP so Codex spends its quota writing code, not finding it.

npx repotracer@latest setup

~63% cheaper (2.68x quota stretch) · ~25% faster · 0 regressions · 0 hallucinated paths
_{Measured on real paired runs (DeepSWE, SWE-bench Astropy, and production bug fixes) with complete scout costs counted. Full writeup · Reproduce it}

_{What the repo_scout MCP tool runs under the hood: plain query in, Luna searches with Read/Glob/Grep, disk-verified file:line citations out in ~35s.}

The problem

Codex subscriptions have a fixed monthly budget. Every time Codex searches your repository, reading files, grepping for symbols, mapping dependencies, it burns that budget on work that writes zero lines of code.

On the tasks we measured, search ate 30-60% of the total cost. That budget could have gone toward actual edits.

RepoTracer moves search to Luna, a model that costs a fraction of Sol, through an MCP tool call. Same code gets written. Your subscription lasts longer.

What you get

Codex calls a tool, not a suggestion. repo_scout(query) is an MCP tool call. Codex can't ignore it, reinterpret it, or go Google how to do it. It calls the function, gets results.

Cheap search, not cheap quality. Each scout runs in its own thread with just the query and read-only access. No conversation history dragged along. Costs a fraction of a full subagent.

No hallucinated paths. Every file path and line range gets checked before Codex sees it. If Luna returns a file that doesn't exist, RepoTracer drops it.

Knows when to do nothing. 42/42 routing decisions correct. If the edit target is already obvious, no scout runs. You don't pay for a search you don't need.

One command. npx repotracer@latest setup. No API keys, no background service, no Rust toolchain. Uses your existing Codex login.

Updates itself. New versions download and verify automatically when Codex starts. Nothing to maintain.

Install

npx repotracer@latest setup

setup downloads the native binary, verifies its SHA-256 checksum, installs it under ~/.repotracer/bin, registers the MCP server, and adds a managed routing block to ~/.codex/AGENTS.md. It reuses your existing Codex login.

No Rust toolchain, second API key, background service, or hand-written config is required.

Need to remove or reconfigure? Run the same command again and pick Uninstall:

npx repotracer@latest setup

RepoTracer is already configured.
Use arrow keys, then Enter. Esc to cancel.

  > Update the configuration
    Uninstall RepoTracer

That removes the MCP entry, the routing block, and the local config. Your Codex login and settings are untouched, and every file it edits is backed up alongside the original first. repotracer uninstall --yes does the same thing without the menu.

Preview the changes:

npx repotracer@latest setup --dry-run

Want the repotracer command in your own shell too?

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

Updating

RepoTracer updates itself. Each time its MCP server starts, it checks the release feed, verifies the new binary's SHA-256 against the published checksums, and replaces the copy in ~/.repotracer/bin. The new version takes effect the next time you start Codex.

It only ever replaces that one binary, then refreshes RepoTracer's managed MCP entry and AGENTS.md block. A cargo install build or source checkout is left alone.

Automatic updates are on by default. Restart Codex after an update for the new binary and managed integration files to take effect. To disable automatic updates, set updates.automatic = false in ~/.repotracer/config.toml or set REPOTRACER_NO_UPDATE=1. You can still update a disabled installation with npx repotracer@latest setup.

To update on the spot:

repotracer update

How it works

You ask Codex to fix something
  Codex decides it needs to find code first
  calls repo_scout(query) via MCP
    RepoTracer spins up an isolated Luna thread
    Luna searches with Read, Glob, Grep
    RepoTracer validates every path and line range
  Codex gets back verified file:line citations
  Codex reads the cited code and makes the edit

The scout runs GPT-5.6 Luna at medium reasoning on the fast service tier. It starts clean, no conversation history, no inherited context, and can only read. It cannot edit, delete, commit, or push.

One tool call. Structured output. Verified results. No prompt interpretation, no web searches for "how to start a Luna agent."

"Can't I just put 'use Luna' in agents.md?"

You can try. We did. Here's what happens.

Sol doesn't have to follow prompt instructions. It interprets them however it wants, or ignores them, or goes and searches the web for "how to start a Luna agent" instead of starting one. An MCP tool call is a function call. Sol calls it, gets back results, moves on.

Subagents also inherit the full conversation context. On a long session that inherited context alone can cost more than the search was supposed to save. RepoTracer starts an isolated thread with only the query and read-only tools. No history.

And raw Luna output is inconsistent. Without validation you get wrong paths, bad line numbers, phantom files. RepoTracer checks every citation before returning it. Bad results get dropped.

We spent weeks benchmarking both approaches. The one-line agents.md instruction consistently cost more than not using it at all. The MCP approach is the one that actually showed up in the numbers.

Benchmarks

Tested on DeepSWE (industry-standard multi-language coding tasks) and MAH-SWE (a benchmark built from real agentic coding sessions on production repositories, not synthetic prompts).

Complete task cost, RepoTracer's usage included.

If a task costs 62.68% less, a fixed budget covers 2.68 of them instead of 1. That's where the 2.7x headline comes from.

Paired runs hold everything constant: same model, same prompt, same repo commit, same timeout. The only difference is whether RepoTracer is installed.

42/42 routing decisions correct. 24/24 holdout. 3/3 real-task runs passed both verifiers. 87 workspace tests passing, formatting and strict Clippy clean.

Every run is published with raw artifacts. Full methodology, evidence, and benchmarks.

Automatic routing across all tasks (zero manual switching)

You never need to turn RepoTracer on or off. Leave it installed and keep prompting Codex normally for every task.

The built-in router (42/42 verified decisions) automatically chooses the optimal route for every prompt:

• Broad exploration & cross-file tasks: Delegates search to Luna via MCP, slashing complete task spend by up to 63%.
• Small or single-file edits: Handles the task directly with Sol with zero scout calls, zero latency penalty, and zero overhead.

You get massive savings on complex tasks with zero penalty on small edits.

Current support

Luna medium is the production default because it hits the optimal performance frontier: 4.00/4.00 (24/24 perfect evaluations) in double-blind quality grading, delivering comprehensive source grounding in ~58s without the latency or token bloat of higher tiers.

CLI

repotracer "where is auth handled?"      # scout the current repository
repotracer scout "trace refresh rotation"
repotracer serve                         # MCP over stdio
repotracer setup
repotracer update                         # replace the binary with the newest release
repotracer doctor
repotracer status
repotracer config --init
repotracer benchmark
repotracer uninstall --yes

--json is available on scout, doctor, and status.

Custom GPT endpoint

repotracer \
  --base-url https://models.example.com/v1 \
  --model gpt-5.6-mini \
  setup

Set REPOTRACER_API_KEY when the endpoint requires authentication.

Security

• Read-only repository access
• Repository-root and symlink-escape checks
• Returned path and line-range validation
• No default telemetry
• Provider credentials remain with the official CLI

See SECURITY.md.

Resources

• repotracer.tech
• Benchmarks
• Architecture
• Why complete-task measurement matters
• Microsoft FastContext paper

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

Develop

cargo test --workspace
cargo run -p repotracer -- doctor
cargo run -p repotracer -- scout "where is config loaded?" --mock