Skip to main content

Cloud Agents

Claude Code, Codex and Pi, running on TOSE

Install a coding agent in your workspace and it gets its own private machine on TOSE with a persistent disk. Work with it in its console, or hand it tasks from ChatGPT, Claude or another MCP client. It can clone your GitHub repositories, push and open pull requests when you allow it, and sleep when it is idle.

Three agents to choose from

  • Claude Code

    Sign-in
    Your Claude plan, with /login in the agent’s own terminal, or an Anthropic API key. TOSE never sees the sign-in.
    Sizes
    1, 2 or 4 vCPU and 2, 4 or 8 GiB RAM.
    Tasks over MCP
    From release 1.1.0, which can also continue a finished task.
  • Codex

    Sign-in
    An OpenAI API key, stored as an encrypted secret of the agent.
    Sizes
    1, 2 or 4 vCPU and 2, 4 or 8 GiB RAM.
    Tasks over MCP
    From release 1.1.0, which can also continue a finished task.
  • Pi

    Sign-in
    Your own key for Anthropic, OpenAI or Google; from release 1.1.0 also OpenCode Go, DeepSeek, Moonshot, OpenRouter, Cloudflare Workers AI, xAI or Z.AI.
    Sizes
    1 or 2 vCPU and 1, 2 or 4 GiB RAM.
    Tasks over MCP
    From release 1.2.0, which can also continue a finished task.

Two ways to give it work

In its console

A workspace owner or admin opens the agent’s terminal from its page in the dashboard. The console has no public address: TOSE checks the person’s sign-in and role on every request and closes it when either ends.

As a task, from another AI

Connect the TOSE MCP server to ChatGPT, Claude or another MCP client, then ask it to give a task in plain language, such as “fix the failing test in this repository”. The agent runs it headless in the same /workspace, and TOSE keeps the record: who gave it, its status, the agent’s final message and the files it changed.

The four MCP tools for tasks

They are part of the TOSE MCP server at https://app.tose.sh/mcp. Each needs its own scope and a workspace owner or admin.

ToolWhat it does
cloud_agent_list
agents:read
Lists the workspace’s agents: release, status, whether they accept tasks, the active task and the agent’s page.
cloud_agent_run_task
agents:execute
Gives an agent a task with a timeout from 1 to 120 minutes (30 by default), or continues a finished one.
cloud_agent_get_task
agents:read
Follows a task: status, events, the final message and the changed files.
cloud_agent_cancel_task
agents:control
Stops a running task.

The tools never open the console, never take a secret and never switch task delegation on. When a task is refused (the agent sleeps, is busy, is not signed in, or the owner has not allowed tasks), the answer names the code and the agent’s page where a person fixes it. Connect the TOSE MCP server

GitHub repositories, commits and pull requests

  • Repositories come from the GitHub App connected to your workspace: no URL, personal access token or SSH key. An agent can have up to 10, each cloned into its own folder under /workspace. A workspace owner or admin attaches, changes or removes them on the agent’s page; a connected app cannot.
  • A checkout is read-only by default: the agent can commit locally but not push. Push is switched on per repository, for every branch or only for branches named tose/<agent>/….
  • The agent never holds a GitHub token. It pushes through the TOSE git proxy, which mints a short-lived token for that one repository only while the agent runs.
  • Managing issues and pull requests is a separate switch per repository. The agent then reaches them through the TOSE GitHub API proxy. Merging, repository contents, settings, hooks and Actions are refused, so a person still reviews and merges.

Auto-sleep

  • Closing the browser does not stop an agent, so each one can sleep when it is idle. Choose Off (the default), 15, 30, 60 or 120 minutes, at install or later on the agent’s page.
  • Idle means no input or output in its terminal and no command still running; a running task keeps it awake. When the agent sleeps, compute stops and the disk, with your files and the provider sign-in, is kept.
  • It wakes only when someone presses Wake on its page or calls the wake route, with a funded wallet. A task sent to a sleeping agent is refused and never wakes it. A conversation in progress does not survive a sleep; files do.

What it costs

  • The agent runs on your prepaid workspace wallet at the standard rate: $0.02 per vCPU-hour and $0.01 per GiB of RAM per hour. 1 vCPU and 2 GiB running all month is $29.20; sleep cuts that for every idle hour. Tasks add no charge of their own.
  • The disk is $0.15 per GiB-month, billed while it exists, also while the agent sleeps or is stopped. /workspace is 5 GiB by default (10, 20 or 50 GiB at install) plus 1 GiB for /home/agent: $0.90 a month at the default. A disk can grow later and never shrinks.
  • TOSE does not bill model usage. Your provider bills it to your own key or plan, so set a spending limit there.
  • When the wallet runs out, compute stops and the disk is kept.

See prices

People stay in control

  • A task is plain language to an agent that runs commands without asking, and it can read and send back anything the agent can, including its sign-in or key. Give agents:execute only to MCP clients you trust that much.
  • Each agent has a switch, Accept tasks from connected apps and API keys, on its page. Only a signed-in owner or admin can turn it on or off; no key or connected app can.
  • A person can always stop the agent, put it to sleep or delete it; stopping or sleeping also cancels its running task. The agent has no TOSE credential, so it cannot reconfigure itself or lock a person out.
  • Every agent runs with a private console, a default-deny network policy for inbound traffic, a non-root user and a read-only root filesystem. It can reach public internet hosts, never private, metadata or cluster addresses.

Limits today

  • One task at a time per agent. For work in parallel, install several agents.
  • An agent installed before its task release (Claude Code and Codex 1.0.0, Pi before 1.2.0) does not take tasks. Releases are not upgraded in place: install the new release as a new agent and move your work over with git.
  • Opening the console and signing in to a provider need a signed-in workspace owner or admin. A connected app cannot attach or change repositories.
  • A task never starts or wakes an agent, and an unfinished task is never replayed after a restart.

Get started

  1. 1. Install an agent

    In app.tose.sh, open your workspace and choose Cloud Agents in the menu. Pick Claude Code, Codex or Pi, its size, disk, auto-sleep and, if you want, a GitHub repository and whether it may push.

  2. 2. Sign it in to its provider

    Add the provider key when you install, or for Claude Code with a Claude plan open the console and run /login.

  3. 3. Connect your AI to TOSE

    Add https://app.tose.sh/mcp as a remote MCP server in ChatGPT, Claude or another client, sign in to TOSE and approve the agent scopes.

  4. 4. Give it a task

    Ask your AI to find the agent with cloud_agent_list, give it a task with cloud_agent_run_task and follow it with cloud_agent_get_task. Review the diff, the test results and the pull request yourself.

app.tose.sh

Documentation