> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tracks.alpic.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Use Tracks from your agent

> Paste one prompt into Claude Code or Cursor to set Tracks up and compare two agents

Your coding agent can drive Tracks for you. It learns how from the Tracks skill, a file at [docs.tracks.alpic.ai/skill.md](https://docs.tracks.alpic.ai/skill.md) that teaches it the `tracks` CLI.

## Before you start

1. In the [web app](https://tracks.alpic.ai), open your profile menu at the bottom of the sidebar, then **API keys**, and create a key.
2. Set it in the terminal you start your agent from:

   ```bash theme={null}
   export TRACKS_API_KEY=trk_...
   ```

Your agent runs the `tracks` CLI, so it needs Node.js. Runs start only while your [balance](https://docs.tracks.alpic.ai/concepts/billing.md) is above \$0.

## Paste this prompt

Replace the task in quotes with one of yours. Runs start from an empty folder unless the experiment picks a repository from your library.

```text theme={null}
Set up Tracks for this project, then use it to compare two coding agents.

1. Download https://docs.tracks.alpic.ai/skill.md and save it as
   .claude/skills/tracks/SKILL.md (in Cursor: .cursor/skills/tracks/SKILL.md).
   Read it and follow it.
2. My API key is in TRACKS_API_KEY. Never print it.
3. Create an experiment that gives Claude Code and Codex this task:
   "Write a Python script that converts a CSV file to JSON, with tests, and run them".
   Write 2 or 3 yes-or-no evaluations that tell whether a run did it well.
   Use 1 repetition, a $1 budget and a 10-minute time limit per run.
4. Launch it with --wait, then tell me each agent's verdict and score,
   and what the execution cost.
```

The agent saves the skill, writes the experiment, launches it and reports back. Open **Experiments** in the web app to see the same results.

## Where the skill goes

**Claude Code**

Save it as `.claude/skills/tracks/SKILL.md` in your project, or in your home directory to use it in every project.

**Cursor**

Save it as `.cursor/skills/tracks/SKILL.md` in your project, or in your home directory to use it in every project.

## Prefer the MCP server

If your agent can't run commands, connect it to the Tracks MCP server instead: it offers the same actions as tools. See [MCP server](https://docs.tracks.alpic.ai/mcp/overview.md) for the setup in Claude Code, Cursor and other clients. Then ask it for the comparison in steps 3 and 4 of the prompt above.

> [!NOTE]
>
> An experiment's repositories, skills, MCP servers and command-line tools come from your library. Add them in the web
> app, under **Library**, before asking your agent to use them.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.