# Connect Claude or ChatGPT

> Add mcp.sandscale.dev as a remote MCP connector in Claude, ChatGPT or any remote MCP client and run sandboxes from chat, with your own sign-in.

Add Sandscale as a connector in a hosted assistant and run sandboxes from chat. It signs in as you, so it can do only what you can.

## Connect

In Claude, ChatGPT or any client that supports remote MCP servers, add a custom connector with this address:

```
https://mcp.sandscale.dev/mcp
```

The client then does the rest itself. It finds Sandscale's sign-in at auth.sandscale.dev, registers, and opens a Sandscale sign-in page. Sign in and approve the app. The approval page names the app and says it is not made by Sandscale. When you approve, the connector is linked to your organization.

Coding agents on your own machine, such as Claude Code, Codex and Cursor, connect to the same address. A local server, [`sandscale mcp`](https://sandscale.dev/docs/reference/mcp), comes with the CLI, which is coming soon.

## What it can do

Start, run, pause, snapshot and stop sandboxes; move files in and out; list, start and stop your workspaces; look at your object storage buckets; and list the names of your Vault secrets. Every tool call goes through the same API, permissions, plan limits and audit log as everything else. A refusal comes back to the assistant with the API's own error code.

Each tool takes a `region`. For now the connector answers for the United States, and another region is refused with `region_unavailable`.

## What it cannot do

- Read Vault values. It lists secret names only.
- Act as anyone else. Each call runs as you, with a short-lived credential that never leaves the server.
- Go unrecorded. Every call is in your organization's audit log: the tool, the region, the outcome and the app, never the arguments.
- Run without limits. Each app can make up to 120 calls a minute, and each person up to 60.

## Disconnect

Remove the connector in the assistant, or revoke it with `DELETE /api/v1/me/mcp-connectors/{id}`. Revoking ends the app's sessions at once. A page for this in the console is coming soon.

## Tools

15 tools.

| Tool | What it does | Changes things |
| --- | --- | --- |
| `sandbox_list` | List your sandboxes in the region with their state. | No |
| `sandbox_create` | Start an isolated sandbox (Firecracker microVM, network off by default). Your plan decides how many boxes and how large. | Yes |
| `sandbox_run` | Run a command in a sandbox and return its exit code, standard output and standard error. | Yes |
| `sandbox_upload_file` | Write a file into a sandbox. Give text in content or bytes in content_base64. | Yes |
| `sandbox_download_file` | Read a file from a sandbox. Text comes back as text, other bytes as base64. | No |
| `sandbox_pause` | Pause a running sandbox (its memory and disk are kept). | Yes |
| `sandbox_resume` | Resume a paused sandbox. | Yes |
| `sandbox_snapshot` | Save the sandbox's /workspace as a snapshot a new sandbox can start from (Pro plan and up). | Yes |
| `sandbox_stop` | End a sandbox. Its disk is deleted; take a snapshot first to keep it. | Yes, cannot be undone |
| `workspace_list` | List your cloud workspaces (persistent cloud computers) with their state. | No |
| `workspace_start` | Start a stopped workspace. | Yes |
| `workspace_stop` | Stop a running workspace (its disk is kept). | Yes |
| `storage_list_buckets` | List your object storage buckets (S3-compatible) with their usage. | No |
| `storage_describe_bucket` | Show a bucket's settings, endpoints and usage. | No |
| `vault_list_secrets` | List the names of secrets in a Vault KV mount. Values are never returned by any Sandscale MCP tool. | No |

This page is generated from the remote MCP server's tool table. Do not edit it by hand.

---
Source: https://sandscale.dev/docs/guides/connect-claude-chatgpt
