---
title: Catalog agents
description: Coding CLIs Redline installs and drives, on your own account.
---

A catalog agent is a headless coding CLI that Redline knows how to install into
the machine image, invoke with your task, and read the output of. You supply the
credential; the model usage is billed to you by that provider.

## The catalog

| Agent | Vendor | Connect with |
|---|---|---|
| **Claude Code** | Anthropic | A Claude subscription (browser approval), or an Anthropic API key |
| **Codex** | OpenAI | An OpenAI API key |
| **Cursor** | Cursor | A Cursor API key |
| **GitHub Copilot** | GitHub | A GitHub token on an account with Copilot |
| **OpenCode** | OpenCode | An OpenRouter key, or an Anthropic key |
| **Grok Build** | xAI | An xAI API key |
| **Gemini CLI** | Google | A Gemini API key |

Each is installed into the run's image and invoked in its own non-interactive
mode with approvals bypassed — nobody is at the keyboard to confirm a tool call,
and the machine is a disposable container.

## Connecting one

On the **Agents** page, press **Connect** on the agent you want. Where an agent
offers more than one method, they are ordered as the vendor prefers them —
subscription before key.

### Pasting a key

Type or paste the key and press Connect. Redline checks its rough shape as you
submit, so a key pasted into the wrong field is caught here rather than twenty
minutes into a run. There is a link beside the field to the vendor's own console
page where a key is created.

### Approving a subscription

Claude Code can run on a Claude subscription instead of an API key. That is not
a pasted secret — it is Anthropic's own approval flow, run from this page:

1. **Press “Approve on claude.ai”**

    A tab opens on Anthropic's approval page.

2. **Approve, and copy the code**

    Anthropic shows you a string that looks like `abc123#xyz789`.

3. **Paste the whole thing back**

    Redline exchanges it for a token pair. The verifier never leaves the server,
    and the code is single-use.

## Enabling is the same act as connecting

Every agent starts switched **off**, and the switch on an unconnected agent
opens the connect panel rather than flipping. An agent with no credential cannot
run anything, and a switch that promises otherwise is a lie.

Only enabled agents are offered by the experiment builder, and a launch that
names an agent this project has not connected is refused with that reason —
before it spends anything, rather than failing on the machine twenty minutes
later.

Turning a connected agent off keeps the credential and simply removes it from
the builder. **Disconnect** removes the credential itself.

## What happens to your key

> **Note**
>
> A key goes in and never comes out. It is encrypted at rest, and what leaves
> the database for the interface is the provider, a hint, and when it was
> connected — never the value. The plaintext travels in exactly one direction:
> into a worker that is about to start a run, as an environment variable inside
> that run's container.

To replace a key, disconnect and connect again. Redline cannot show you the old
one, by design.

## Cost

A run's cost and token count come from what the agent itself reported and are
shown on the run, the experiment and the home page. They are a record of what
you spent on your own provider account — Redline does not bill for model usage
and does not proxy it.