---
title: "Delegate tasks to a Cloud Agent from Claude, ChatGPT or Claude Code via MCP"
description: "Connect the TOSE MCP server to Claude, ChatGPT, Claude Code, Cursor or VS Code, then assign and track Cloud Agent tasks with the cloud_agent_* tools."
url: https://tose.sh/en/blog/delegate-tasks-to-cloud-agents-via-mcp
lang: en
category: "Guides"
tags: ["cloud-agents", "mcp", "claude", "chatgpt", "guide"]
author: "TOSE"
published: 2026-11-13T12:49:41.779384Z
updated: 2026-11-13T12:49:41.779449Z
---

# Delegate tasks to a Cloud Agent from Claude, ChatGPT or Claude Code via MCP

**TL;DR:** Copy the MCP server URL from workspace Settings, Agent access, and add it to Claude, ChatGPT, Claude Code, Cursor or VS Code. Four tools: cloud_agent_list, cloud_agent_run_task, cloud_agent_get_task, cloud_agent_cancel_task. One task per agent at a time, 1 to 120 minutes (30 by default); a task never wakes a sleeping agent.

> **info**
> **Cloud Agents on TOSE series** · Part 3 of 5
> 1. [Introduction](/en/blog/introducing-cloud-agents-on-tose) · 2. [Create your first Cloud Agent](/en/blog/create-your-first-cloud-agent) · 3. **Delegate tasks via MCP** (this article) · 4. [GitHub and pull requests](/en/blog/cloud-agents-github-pull-requests) · 5. [Pricing and auto-sleep](/en/blog/cloud-agent-pricing-and-auto-sleep)

[Video: Delegate tasks to a Cloud Agent via MCP](https://cdn.tose.sh/blog/d5a941a6d8cdb5d4ccf830ae16f6027a.mp4)

![A chat that delegates a task to a Cloud Agent over MCP](https://cdn.tose.sh/blog/50d926e72e0328c605ad94cd2e9e8ae2.webp)

*Mockup of a neutral chat interface with sample data.*

## Requirements for accepting tasks

• The agent was installed from a release with task support: **Claude Code 1.1.0, Codex 1.1.0 or Pi 1.2.0**. Older releases return `TASKS_UNSUPPORTED`.

• In the agent's **Settings** tab, **"Accept tasks from connected apps and API keys"** is on. It is on by default for new installs; when off, tasks are refused with `TASKS_DISABLED`.

• The agent is **awake**. A task never wakes a sleeping agent (`AGENT_SLEEPING`).

• The agent is signed in to its provider; otherwise the task returns `PROVIDER_SIGN_IN_REQUIRED`.

No agent yet? Follow [part 2](/en/blog/create-your-first-cloud-agent).

## Connect the TOSE MCP server

Open the workspace **Settings**, go to **Agent access** and copy the **MCP server URL** (usually `https://app.tose.sh/mcp`). Then add it to your client:

| Client | How to add it |
| --- | --- |
| Claude | Settings → Connectors → Add custom connector → paste the URL |
| ChatGPT | Settings → Apps & Connectors → Create (developer mode) → paste the URL, choose OAuth |
| Claude Code | `claude mcp add --transport http tose <MCP server URL>` |
| Cursor | In `~/.cursor/mcp.json`: `{"mcpServers":{"tose":{"url":"<MCP server URL>"}}}` |
| VS Code | Run *MCP: Add Server*, choose HTTP, paste the URL |

*Open the workspace Settings, go to Agent access and copy the MCP server URL (usually `https//app.tose.sh/mcp`). Then add it to your client*

When the client asks, sign in to TOSE, pick the workspace and choose what to allow:

• `agents:read`: list agents and read tasks;

• `agents:execute`: assign tasks;

• `agents:control`: cancel tasks.

The person granting access must be a workspace owner or admin. Note the warning on `agents:execute`: a task can read and send back anything the agent can, including its sign-in. Grant it only to applications you trust.

## The four MCP tools

| Tool | What it does |
| --- | --- |
| `cloud_agent_list` | Lists agents with status, release, task support and any running task |
| `cloud_agent_run_task` | Assigns a task to an agent |
| `cloud_agent_get_task` | Reads a task's status and events; each call can wait up to 25 seconds |
| `cloud_agent_cancel_task` | Stops a running task |

*The four MCP tools*

You rarely call these directly. Describe the job in plain language and the AI picks the right tool.

## The life of a task

![The life cycle of a task](https://cdn.tose.sh/blog/750ca300c9c733a3087634f494eb53fe.webp)

A task starts as `starting`, moves to `running`, and ends as `succeeded`, `failed`, `cancelled`, `timed_out` or `lost`. The result includes:

• `final_message`: the agent's closing summary;

• `changed_files`: changed files by repository and status;

• `problem_code` and `ended_by` when something went wrong or the task was stopped.

## Limits to keep in mind

• **One task per agent at a time, with no queue.** A new task while the agent is busy returns `AGENT_BUSY`. Use several agents to work in parallel.

• **Run time** is set with `timeout_minutes`, from 1 to 120 minutes, 30 by default.

• **Prompts** can be up to 100,000 bytes.

• **Continue earlier work** by passing `continue_task_id` of a finished task, so the agent picks up in the same context.

• **Empty wallet:** tasks are refused with `INSUFFICIENT_BALANCE`.

## Follow tasks on the agent page

Every task, wherever it came from, appears in the **Tasks** tab in three columns: running, finished, and failed or stopped.

![The agent's task board](https://cdn.tose.sh/blog/a3198c546008dcfb05975d3040ca2abe.webp)

*Mockup of the Tasks tab with sample data.*

## A brief that lets the agent work alone

```
Context: web-app repository; a refresh-token bug signs users out after 15 minutes.

You may change: src/auth/ and tests/auth/.
Do not change: deploy config, migrations, package.json.

Done when:
- a test reproduces the bug, failing before the fix and passing after it;
- all existing tests still pass;
- you report the cause, the files changed and how to verify.
```

A good brief has three parts: context, scope and a definition of done. Leave one out and the agent has to guess it.

## When the result comes back

"Succeeded" does not mean correct. Check `changed_files`, read the diff and rerun the tests. If the agent may push, the result lands on a branch or a pull request for review, as the next part explains.

**Previous:** [← Create your first Cloud Agent](/en/blog/create-your-first-cloud-agent) **Next:** [Connect GitHub and receive pull requests →](/en/blog/cloud-agents-github-pull-requests)

## FAQ

### Which agents accept tasks over MCP?

Agents installed from Claude Code 1.1.0, Codex 1.1.0 or Pi 1.2.0 that are awake, have task intake turned on and are signed in to their provider.

### Can one agent take several tasks at once?

No. Each agent runs one task with no queue; a new task while it is busy returns AGENT_BUSY.

### Which permissions does delegating over MCP need?

agents:execute to assign tasks, agents:read to list agents and read tasks, agents:control to cancel. The person granting access must be a workspace owner or admin.
