> For the complete documentation index, see [llms.txt](https://docs.computesdk.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.computesdk.com/platform/mcp.md).

# MCP server

Connect Claude, ChatGPT and other AI assistants to the ComputeSDK Platform through its remote MCP server: run code in cloud sandboxes, work with Actions CI, check provider keys and read benchmark resu

The ComputeSDK Platform has a remote [Model Context Protocol](https://modelcontextprotocol.io) server. Connect it to an AI assistant and the assistant can create cloud sandboxes, run commands and long-running processes, read and write files, share a preview URL for a server it starts, work with your org's Actions CI runs, show which provider keys are connected, and read benchmark results.

**Server URL:** `https://platform.computesdk.com/mcp` (streamable HTTP)

Nothing runs on your computer: every tool call goes to this server, and commands run inside cloud sandboxes.

## Connect

### Claude

* **Claude (web and desktop):** add ComputeSDK Router from the connectors directory, or go to **Settings → Connectors → Add custom connector** and enter the server URL.
* **Claude Code:** install the plugin, which adds the server and skills that help Claude use the tools well:

  ```
  /plugin marketplace add computesdk/computesdk
  /plugin install computesdk-router@computesdk
  ```

  Or add only the server: `claude mcp add --transport http computesdk https://platform.computesdk.com/mcp`

### ChatGPT

Install the **ComputeSDK Router** app from the ChatGPT apps directory. In developer mode you can also add the server URL as a custom MCP app.

### Other MCP clients

Any client that supports remote MCP servers with OAuth works: point it at the server URL. Clients that can send headers can use an org API key instead of OAuth (`Authorization: Bearer <key>`); create one under **Settings → API keys** on the platform.

## Sign-in and permissions

The first time the assistant uses a tool, a ComputeSDK sign-in opens in your browser:

1. Sign in to your ComputeSDK account.
2. Choose the organization the connection uses. Usage is billed to that org.
3. Approve the permissions:

| Permission       | Scope             | Allows                                                                    |
| ---------------- | ----------------- | ------------------------------------------------------------------------- |
| Read sandboxes   | `sandboxes:read`  | List and inspect your sandboxes, files, processes and URLs                |
| Use sandboxes    | `sandboxes:write` | Create and destroy sandboxes, run commands, write files, manage processes |
| Read Actions     | `actions:read`    | Read CI runs, jobs, logs and settings                                     |
| Manage Actions   | `actions:write`   | Dispatch, cancel and re-run CI runs; change Actions settings              |
| Read providers   | `providers:read`  | Provider key status and rates                                             |
| Manage providers | `providers:write` | Verify or remove provider keys, set reporting rates                       |
| Read benchmarks  | `benchmarks:read` | Benchmark runs and results                                                |
| Stay signed in   | `offline_access`  | Keep the connection after you close the browser                           |

Settings changes, provider keys and repo controls also need you to be an org owner or admin. To disconnect, remove the app in **Settings → Connected apps** on the platform, or remove the connector in your assistant.

## What the assistant can do

**Sandboxes**

* See what you can run on and what each size costs (`sandboxes_get_options`), and price one create (`get_quote`).
* Create a sandbox (`create_sandbox`), on the router's choice or a provider you name, on your own keys or ComputeSDK's managed accounts.
* Run commands (`run_command`), start long-running processes (`start_process`, `process_status`, `wait_process`, `kill_process`, `write_process_stdin`), and work with files (`write_file`, `read_file`, `make_directory`, `remove_path`).
* Get a public preview URL for a port a server listens on (`sandbox_url`).
* List, inspect and destroy sandboxes (`list_sandboxes`, `get_sandbox`, `destroy_sandbox`), and see spend (`sandboxes_get_costs`).

**Actions (CI)**, for repos connected through the ComputeSDK GitHub App:

* List repos, branches, workflows, runs and the schedule; read a run, its jobs and step logs (logs are secret-redacted).
* Dispatch, cancel and re-run runs (any org member).
* Poll a repo, and change repo, workflow and Actions settings (owners and admins).

**Providers**

* List the providers your org can bring its own key for, with key status and a masked hint. Key values are never returned.

**Benchmarks**

* Read benchmarks, runs, per-provider results, reliability and worker logs.

Billing, top-ups and org membership are managed on the platform, not through assistants.

## Good to know

* **Sandbox visibility:** an assistant sees only sandboxes you created through a connected app. Other members' sandboxes, and ones made in the dashboard, return "Sandbox not found".
* **Billing:** sandboxes on managed accounts bill your org's prepaid credit while they run; sandboxes on your own provider keys are billed by that provider. Sandboxes expire on their own (30 minutes by default); destroying them sooner stops billing.
* **Preview URLs:** a server must listen on `0.0.0.0` and the exact port requested.
* **Response size:** assistant connections get compact results (for example, the newest runs and only their failing jobs). Ask for one run, or narrow with filters, for the full detail.
* **Missing permission:** if a tool answers `insufficient_scope`, the connection predates that permission. Disconnect and connect again to approve it.

## Data and privacy

Tool calls carry only what the assistant sends for that action, such as a command, a file, or a run id. Sandboxes run on third-party compute providers, and code running in a sandbox can reach the internet. See the [privacy policy](https://www.computesdk.com/privacy) for what ComputeSDK collects, how long it keeps it and your choices. Questions: [support](https://www.computesdk.com/support).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.computesdk.com/platform/mcp.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
