Agents

Metaphor coordinates the coding agents you already use, including Claude Code and Codex.

For each task, you choose the repository and agent. Metaphor gives the task an isolated worktree, runs the agent there, tracks its changes, and verifies the result before the work moves forward.

Supported agents

Metaphor currently supports:

• Claude Code
• Codex

Both run using the agent CLI and subscription you already have.

How agents work in Metaphor

When you start a task:

1. You choose the repository and agent.
2. Metaphor creates an isolated worktree for the task.
3. The agent starts inside that worktree with your request and relevant context.
4. Metaphor tracks the agent session and resulting changes.
5. Metaphor verifies the result before the work moves forward.

The agent works on the code, while Metaphor coordinates the workflow around it.

Metaphor follow-ups

Metaphor doesn’t have to stop at the agent’s first result.

During verification, Metaphor checks the work against the task’s checks and can test the resulting changes. If something doesn’t pass, Metaphor sends the relevant context back to the agent as a follow-up so it can continue working.

This creates a loop:

Agent works → Metaphor verifies → Follow-up → Agent fixes → Verify again

The task moves forward once the result passes verification.

Learn more about Checks.

Continue with another agent

You can switch agents without starting the task over.

The next agent continues from the task’s existing worktree and the context Metaphor has gathered. The work stays with the task, not with a particular agent.

This lets you use different agents at different stages of the same task without manually transferring changes or starting over.

Pull request feedback

Metaphor can also send relevant pull request feedback back to an agent.

When review tools like CodeRabbit leave comments on a pull request, Metaphor can use that feedback to continue the task and update the changes.

Approvals

Some agents require approval before running certain commands.

Codex can be configured to run without waiting for command approval when you want a task to run unattended. This is useful for longer-running tasks or tasks started remotely, such as from Slack.

Configure approval behavior in Settings.

Bring your own agent

Metaphor doesn’t replace your coding agents with a hosted model.

You continue using the agent CLIs and subscriptions you already have. Metaphor manages the workflow around them - worktrees, task context, verification, Git operations, and coordination between agents.