llms.txt
Compare with other sites' llms.txt files →# atgent
> Email your agent a task. Codex or Claude Code runs it on your own computer and replies in the same Gmail thread. Free.
Page: https://atgent.dev
Package: https://pypi.org/project/atgent/
Publisher: Compound Labs (https://thecompound.tech)
## What it does
You email your agent a task. Codex, Claude Code or another coding agent runs it on your own computer, on its own branch, and replies in the same Gmail thread.
Use the Gmail you already have. Add +agents to your own address. Gmail delivers it to your normal inbox, so you need no new email account.
You set up a small runner on your computer once. You need no phone app, SSH, tunnel, hosted mail service or agent plugin.
The runner reads your Gmail Sent folder, not your Inbox. Mail from other people lands in your Inbox, so their mail cannot start a task.
## Writing tasks
The subject line picks the agent. Settings follow in any order. atgent parses the subject. Put the agent first, then add a project, a model or other settings in any order. You do not need a colon or brackets. Write the task in the email body.
- codex. Codex runs the task in your default project, on its own branch.
- claude website opus. Claude Code runs the task in the project named website with the opus model. Replies in the thread keep that model.
- claude readonly. Claude Code reviews the code and changes no files.
- codex high 30m pr. Codex works with high effort for up to 30 minutes and opens a pull request when it finishes.
- codex terminal. Codex opens the task in a new Terminal window.
- claude opus: fix the login page. Text after a colon is a title for your own inbox. atgent does not read it or send it to the agent.
- Capital letters do not matter. A word that is not a setting does not stop the task. The Queued reply names each word it ignored. The older Claude/website [opus]: form still works.
- A reply with your next instruction. The same session continues with your reply.
- A reply of status. You get the status of each turn by email.
- A reply of stop. Queued turns are cancelled and the running agent stops. File changes it already made are kept.
Each update arrives as a reply in your task’s thread. The result is the agent’s final message as formatted email. Every reply ends with the task id and the command to open it in your terminal.
## Forward it, say it, schedule it.
- Forward a bug report or an error alert with a note such as fix this. The forwarded message becomes the task, marked as content from someone else.
- Attach a voice memo. It is transcribed on your computer before the task starts, and the audio does not leave your computer.
- When the agent needs a decision, it ends its turn with a question. Your reply answers it in the same session.
- Put two agents in the subject and both work on the task, each on its own branch. Reply keep codex to keep one.
- Add when to run it to the subject. Each run starts a fresh task, and its result arrives in the same thread.
- Reply save as review to save a task's settings and first instruction. Then send review as the subject.
- A failed GitHub Actions run, a labelled GitHub issue or a webhook starts a task on your computer.
- One email a day lists what ran, what is waiting on you and what failed.
- Gemini CLI, opencode, aider and Cursor's agent also work. Each is one word in the subject.
- In a git project, each task runs on its own branch until you reply merge.
- Result emails list the changed files, a short diff and the check results.
## Continue in your terminal
atgent list shows each email task with its agent, model and status. Select one and press Enter to resume the same session in your terminal. The agent then answers there, not by email. atgent open TASK_ID opens one task directly.
## Sessions (atgent 0.5.0)
Session attachment, MCP, the email directory and optional native session registration are available in atgent 0.5.0. Upgrade with uv tool upgrade atgent, then run atgent start to refresh the compatible runner. CLI, MCP and email use do not require dashboard login.
Connect a persisted Claude Code or Codex session to atgent. With the MCP server and native identity hooks enabled, ask "email me this session." Close the local session, run atgent handoff TASK_ID, and reply from your phone to continue.
Email sessions to your +agents address to get a list of your sessions. Reply open TASK_ID to receive that session's thread. You need no dashboard login.
The optional dashboard is open. Its pairing command, atgent login, is included in atgent 0.5.0. It lists sessions from your paired computers and sends a magic link to the email in your config. Dashboard login is optional. Your CLI and email workflow keep working without it.
The dashboard stores session titles, project aliases, agent and model, task IDs, status and timestamps. It never receives transcripts, prompts, results, attachments, file paths, native session IDs or mailbox credentials.
## Setup
1. uv tool install atgent
2. atgent init. The guided setup asks for your Gmail address, how the runner signs in, your project folders, which installed agents to use and a default model for each. After successful checks it recommends automatic startup. Installing the CLI alone does not start the runner.
3. Turn on IMAP in your Gmail settings. Create a Google app password. Keep it in your OS keychain or password manager and retrieve it with a command. Login services do not inherit Terminal environment variables.
4. atgent doctor checks Gmail, installed agents, runner liveness and automatic startup.
5. atgent install-service enables automatic startup if guided setup has not already installed it. atgent service-status checks the service. Wait for "Mailbox intake armed; runner ready" in state/service.log beside your config before sending the first task; the first successful poll skips old Sent mail. atgent run is a foreground alternative.
6. atgent model claude opus changes an agent’s default model later.
7. atgent mode terminal claude makes Claude tasks open in a new Terminal window by default. atgent mode headless goes back to background runs. atgent remote-control on starts Claude terminal tasks with Remote Control.
## You need
- A Gmail account with IMAP access and an app password or your own OAuth client ID.
- macOS or Linux.
- Python 3.11 or later, and uv.
- Codex or Claude Code, installed and signed in.
- A computer that stays awake and online. The recommended login service starts after you sign in and restarts the runner after failure.
## Docs
- Getting started: https://atgent.dev/docs
- Gmail sign-in, including your own OAuth client: https://atgent.dev/docs/gmail-sign-in
- Projects and config: https://atgent.dev/docs/config
- Continue a coding task by email: https://atgent.dev/docs/continue-coding-task-by-email
- Approval emails: https://atgent.dev/docs/approvals
- How it works: https://atgent.dev/docs/how-it-works
- Connect a session: https://atgent.dev/docs/connect-a-session
- Find sessions by email: https://atgent.dev/docs/session-directory
- Optional dashboard: https://atgent.dev/docs/dashboard
- Optional campaign measurement: https://atgent.dev/docs/campaign-measurement
- Troubleshooting: https://atgent.dev/docs/troubleshooting
- Changelog: https://atgent.dev/changelog
- Support: https://atgent.dev/support
- Privacy: https://atgent.dev/privacy
- Security: https://atgent.dev/security
## Compare
- All tools in one table: https://atgent.dev/compare
- atgent and Claude Code Remote Control: https://atgent.dev/vs/claude-code-remote-control
- atgent and Codex on mobile: https://atgent.dev/vs/codex-mobile
- atgent and Happy: https://atgent.dev/vs/happy
## Price
Free. Your existing Codex or Claude Code plan and its limits apply.