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

# Chat

> Claude or Codex running inside an isolated git worktree per session, so destructive changes never touch your main tree.

The **Chat** tab is where you collaborate with AI agents — Claude (Anthropic) or Codex (OpenAI) — inside Tempo. Every session runs in its own **git worktree**, an isolated copy of your repo on a throwaway branch, so the agent can make destructive changes without ever touching your main working tree.

Open the chat panel from the **Chat** icon in the left sidebar, or use `Cmd/Ctrl+L` from anywhere.

<Frame>
  <img src="https://mintcdn.com/tempolabs/mhsVPkJA5sDhkLjO/images/product/chat/panel.png?fit=max&auto=format&n=mhsVPkJA5sDhkLjO&q=85&s=3694231721d76fb787ce6622cd002b36" alt="The Chat panel" width="1200" height="1600" data-path="images/product/chat/panel.png" />
</Frame>

## Worktrees: isolation by default

When you start a new chat, Tempo creates a fresh worktree on a throwaway branch. The agent works in that copy. Your main branch stays untouched until you explicitly merge.

* **Happy with the work?** Use the share/merge controls at the top of the chat to bring the changes back into your main branch (or open a PR).
* **Not happy?** Discard the worktree and nothing is lost. Your main branch never saw the changes.

This means you can run **multiple agents in parallel** on the same repo — one per worktree — without conflicts.

<Frame>
  <img src="https://mintcdn.com/tempolabs/mhsVPkJA5sDhkLjO/images/product/chat/parallel-worktrees.gif?s=857ee3f20adb356ed17399b21f5ba98c" alt="Two agents working in parallel worktrees" width="1200" height="675" data-path="images/product/chat/parallel-worktrees.gif" />
</Frame>

## Models

Pick a model from the dropdown at the bottom of the chat:

* **Claude** (Opus, Sonnet, Haiku) — sign in via `claude /login` in the terminal.
* **Codex** (GPT-5.4 Codex, GPT-5.4-mini) — sign in via `codex login`.

See the [Quick Start](/guides/quick-start#3-connect-codex-or-claude) for sign-in details.

## Context the agent sees

Every turn, the agent automatically has access to:

* The current worktree (full file tree, git history).
* Any [docs](/product/docs) you've linked to the session.
* Any [canvases](/product/canvas) you've linked to the session.
* Any [issues](/product/issues) you've linked to the session.

This is the unified context layer — instead of pasting requirements into the prompt, you link the PRD and the agent reads it.

## Plan mode

Before letting the agent execute, you can ask it to produce a plan: a structured proposal of what it intends to change. Review and edit before approving. The agent only writes files once you accept the plan.