Rig

Did That Change to Your Agent Help?

Rig is a free, open-source app that runs the same coding tasks with and without a skill, hook, model or CLAUDE.md rule, each in its own sandbox, and shows what really changed, with a 95% interval.

Get Rig on GitHub See Two Real Runs

MIT licence · runs on your machine · free on a local model

Real Run

10 coding tasks, each run with and without the skill price-tags-style on Qwen3.5 9B through Claude Code, locally and for free.

Without the Skill2 of 10 passed
With the Skill9 of 10 passed
+70 pts 95% interval +14 to +91 Helps

Drag left to count fewer tasks and watch the interval widen.

Rig 78Works, but the lift could be as small as +14.

How It Works

Rig runs every task twice, once with the change and once without, and compares the pairs.

  1. Pick the Change

    A skill, a hook, an MCP server, a CLAUDE.md rule, a model, a workflow step, or a change to your code.

  2. Run It Both Ways

    Each task runs with and without the change, each time in its own Docker sandbox. Checks run with no network.

  3. Read the Result

    Rig reports the difference with a 95% interval and a Rig score from 0 to 100. When the evidence is thin, it says so.

Rig's Build page: a seven-step workflow on a canvas, from Clarify the Task to Write the PR, with each step's model and an estimate before any run.
In the app, Build: a workflow's steps, models, skills and rules, each a part you can swap.

Two Real Runs

Both ran for free on Qwen3.5 9B through Claude Code. One skill clearly helped; the other may have cost tokens, and Rig said it couldn't tell.

A House-Style Skill for Price Labels

10 tasks, once each way · scored on the pass rate · 27 October 2026

  • 9 of 10passed with the skill; 2 of 10 without
  • +70 pts95% interval +14 to +91
  • Rig 78works, but the lift needs work
The first run in Rig: ten tasks with and without the skill, 9 of 10 passed with it and 2 of 10 without, +70 pts with a 95% interval of +14 to +91, and a Rig score of 78.
In the app: each task both ways, the interval, and how the score was made.

A Skill Meant to Save Tokens

40 pairs (10 tasks × 4) · scored on tokens per solved task · 28 October 2026

  • 40 of 40passed with the skill; 37 of 40 without
  • +4%tokens per solved task; interval −17% to +27%
  • Rig 17not proven, so no saving is credited
The token run in Rig: 40 of 40 passed with the skill and 37 of 40 without, tokens per solved task 70.9k with against 68.4k without, and a Rig score of 17.
In the app: the saving, its pass-rate guard, and why the score stays low.

Every number above comes from two files in the repo: recorded.json and tokens/recorded.json.

What You Can Measure

Each one is a single part of the agent, swapped in and out on the same tasks.

Install

Rig is a Python engine that serves a web app on 127.0.0.1, so nothing is hosted. You need Python 3.12 or newer, uv, Node 22 or newer, git, and Docker with Compose v2.

It works with Claude Code, Codex, OpenCode and Gemini. A paid run always shows an estimate first and stops at the budget you set.

Install and Check

git clone https://github.com/GokulMV/rig
cd rig/engine && uv sync
(cd ../web && npm ci && npm run build)
uv run rig doctor

Start It on Your Repo

uv run rig serve --root ~/code/my-app
# then open http://127.0.0.1:7357