Skip to the content
how it works

A task in. A branch out.

Cloudfield runs coding agents on a Linux server you own. Each task gets its own container and its own branch. Follow it from any client, answer it when it needs you, and bring the branch back when it is ready.

the server and its clients

One server, and a client wherever you are

The server is a Linux machine with Docker and git. It holds your repositories, runs every agent in a container of its own, and keeps going when your laptop is closed.

  1. The server runs every agent, each in a container of its own.
  2. Clients attach and leave. The server keeps going.
  3. Your phone can tell you when a session needs you.
clients

You link a machine to it once

The first person to link a fresh server becomes its super admin: on a tailnet Tailscale says who you are, elsewhere your SSH key does, or you sign in with GitHub through Cloudfield Edge, which needs no open port on the server. Teammates join after that, by invitation or an admin's approval.

Linking from your machine

  1. The CLI

    On Linux, macOS or Windows. Launch sessions, attach to their terminals, message agents, pull branches and close sessions from a shell.

  2. The desktop app

    The dashboard in its own window, with a notification when a session needs you. The CLI installs it: on Windows it downloads the release's prebuilt app, on Linux and macOS it builds the app on your machine. The install scripts can do this for you.

  3. A browser tab

    The CLI serves the same dashboard from your machine, so the page can pull a session's branch into your clone.

  4. Your phone

    On a tailnet the server can host the dashboard for any device, and that dashboard in your phone's browser can tell you when a session needs you. The Android app opens it full screen over Tailscale or Cloudfield Edge, and notifies you through a UnifiedPush app such as ntfy.

launch

Type a task, pick an agent, launch

A session is one agent on one task, in a container on the server with its own git checkout and its own branch.

Read the quickstartSee it in the app

  1. Type the task and pick an agent.
  2. Launch starts it in its own container.
  3. The agent reads the code and works on the session's own branch.
sessions and modes

Every task runs in its own container

Your home on the server is mounted into the session's container, so the agent has its login and your settings. You pick the agent and a mode when you launch the session.

Enhanced sessions

nik's session in Enhanced mode: Chat shows Claude Code's plan, thoughts and tool calls.

  • Standard opens the agent's own terminal, as it runs on your machine.
  • Enhanced drives the agent over the Agent Client Protocol and shows its plan, messages, tool calls and diffs as controls of their own.
  • Background runs the task once with no interaction, then commits, publishes the branch and finishes.
  • In every mode the agent works without permission prompts by default. The container is the boundary.
session states

The board shows which sessions need you

The board sorts every session into Needs you, Working and Finished. A session moves to Needs you only when a person has to act, and its card says why. Waiting on CI or on its own agents still counts as Working.

Session states and recovery

The northwind-api board. nik's session is Working.

  • Needs you: the agent asked a question or reported it is blocked. Reply from the card, or open the session.
  • Finished, for review: the agent reported success. The session stays open, so a message keeps it going.
  • Mark complete closes the session as completed. Close records it as merged or abandoned.
  • A closed Standard or Enhanced session keeps its container for seven days by default and can be reopened. Its branch stays.
keeps going

Close your laptop and it keeps going

The server supervises the session's container and owns its terminal, so closing a client does not stop the session. Reconnect from the CLI or the dashboard to continue watching or steering it.

Read the quickstart

  1. Close your laptop.
  2. The session keeps going on your server.
  3. Reconnect to keep watching or steering it.
review

Review it like any other change

When the agent reports it is done, the session waits for you. Read the diff in the app, send review comments back to the agent, then push the branch and open a pull request, or pull it into your clone.

Reviewing and publishingSee it in the app

  1. Read the session's branch against main.
  2. Comment on a line and send it to the agent.
  3. Push the branch and open a pull request, or pull it into your clone.
review and land

Review the branch, then land it your way

Changes shows the session's branch against the base. Comment on lines and send the comments to the agent as one message. Publish commits, pushes the branch and opens a pull request with your GitHub login, and the session then shows the pull request's checks. Cloudfield never merges a session on its own.

Reviewing and publishing

The session is finished for review, and its pull request #412 has merged on GitHub.

  • Pull fetches the session's branch into your clone from the desktop app, the local dashboard or the CLI. It never switches your checkout.
  • A failing check or a review comment on the pull request can go to the agent as a message.
  • When the pull request merges, close the session as merged.
across repos

One session, several repos

One session can work across several repos. Add them when you launch it, and the agent changes them together, with a branch in each.

Working across reposSee it in the app

  1. One session forks into a lane per repo.
  2. Each change lands in all of them together.
  3. A branch in each. No merge.
agents

Keep your own agent

An agent definition tells Cloudfield how to start an agent's CLI in each mode and where its login lives. Cloudfield ships definitions for Claude Code, Codex and many more agent CLIs, which the guide lists, and you can register a definition of your own.

Agents and custom definitions

  1. A definition says how to start an agent's CLI in each mode.
  2. And where its login lives.
  3. Register a definition of your own.
shipped agents

Installed only when you ask

Cloudfield installs an agent only when you ask, into your environment on the server.

Agents and custom definitions

  • Claude Code

    --agent claude

    Enhanced through an adapter. Login in ~/.claude/.credentials.json

  • Codex

    --agent codex

    Enhanced through an adapter. Login in ~/.codex/auth.json

  • pi

    --agent pi

    Enhanced through an adapter. Login in ~/.pi/agent/auth.json

  • oh-my-pi

    --agent omp

    Enhanced built into the CLI. Login in ~/.omp/agent/

  • OpenCode

    --agent opencode

    Enhanced built into the CLI. Login in ~/.local/share/opencode/auth.json

  • GitHub Copilot CLI

    --agent copilot

    Enhanced built into the CLI. Login in ~/.copilot/config.json

  • Gemini CLI

    --agent gemini

    Enhanced built into the CLI. Login in ~/.gemini/oauth_creds.json, ~/.gemini/.env

  • Cursor CLI

    --agent cursor

    Enhanced built into the CLI. Login in ~/.config/cursor/auth.json

  • Antigravity CLI

    --agent antigravity

    No Enhanced mode. Login in ~/.gemini/antigravity-cli/antigravity-oauth-token

The list is not closed. Any other CLI can be a session's agent: register it as a custom agent with aether agent add.

Each logo is its maker's trademark, shown only to name an agent Cloudfield can launch. Claude is a trademark of Anthropic, the OpenAI mark is a trademark of OpenAI, and Gemini CLI and Antigravity are trademarks of Google LLC. Cloudfield is not affiliated with these companies or endorsed by them.

logins

Keep logins on the server

You sign your agents in from your environment terminal, and their logins stay in your member home on the server.

The environment home

from your phone

Steer a session from a chat

Add Cloudfield as a connector in ChatGPT or Claude, the phone apps included. The connector acts as you, under the same role checks as the app.

ChatGPT and Claude connectorsSee it in the app

  1. Ask what a session is doing.
  2. Tell it what to change. The app asks before it messages the session.
  3. The message reaches the agent at once.
the dashboard

Every session in one window

The desktop app, the local dashboard and the dashboard your server hosts are the same app: a workspace switcher over its sessions, the Board and its Map, and for every session its chat or terminal, its changes, the files in its checkout and its browser.

Watching a session

The Cloudfield desktop app on the northwind-api workspace. The sidebar lists the sessions under Needs you, Working and Finished, with a session's three subsessions nested under it. The Board has the same three columns: a Codex session in Needs you shows its agent's question, two Working sessions warn of a file overlap, a session with subsessions is one card, and Finished holds completed, failed and merged sessions.
Sessions sorted by what needs you. A card says why it is waiting, such as the agent's question, and warns when two sessions edit the same file. A session and its subsessions are one card.
get running

Start with one session

The server needs Linux, Docker and git. Install Cloudfield, then follow the quickstart to connect a client, prepare an agent and review your first branch.

Open the quickstartInstall