> ## Documentation Index
> Fetch the complete documentation index at: https://docs.slipway.sh/llms.txt
> Use this file to discover all available pages before exploring further.

# AI agents

> Point Claude, Cursor, or your own agent at slipway's MCP server and it drives the whole environment lifecycle.

Slipway is built for developers and their agents. The same environment your team reviews in, an AI coding agent can bring up, test against, and tear down on its own, over a first-party [Model Context Protocol](https://modelcontextprotocol.io) server.

## Why it matters

An agent that writes code needs somewhere real to run it. Local sandboxes lie: a different kernel, a different libc, half your environment, none of your data. The bug only shows up after merge.

Slipway hands the agent a real, isolated, internet-reachable environment, so it can close the loop itself: make a change, deploy it, hit the URL, read the logs, and verify the fix before it calls the task done. No tunnels, no local setup, no waiting on a shared staging queue.

## What an agent can do

Over the MCP server, an agent drives the same lifecycle a human does:

* **Spin up** an isolated environment from any branch or PR, and get a public URL back to hit.
* **Inspect** environments, deployments, and repositories, and read a repo's resolved spec.
* **Debug** by streaming logs and running commands inside a running service (for example, run your tests).
* **Check capacity** before it brings things up, so it doesn't hit your plan's concurrency cap.
* **Redeploy, cancel, and tear down** the whole environment when the task is done.
* **Configure** variables and secrets, when its token has the role for it.

An agent isn't a second-class path. It drives the exact environments your team uses, through the same build, deploy, and teardown flow.

## Set it up

<Card title="Connect an agent" icon="plug" href="/configuration/mcp">
  Create an org-scoped token, point your agent at the endpoint, and you're driving environments in two steps.
</Card>

Setup is quick: create a token under **Settings → MCP** in the console, then point your agent at `https://<your-slipway-host>/mcp` with an `Authorization: Bearer` header. For **Claude Code**:

```bash theme={"system"}
claude mcp add --transport http slipway https://<your-slipway-host>/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"
```

The full connection steps for Cursor and Claude Desktop, the complete tool list, and the role table are on the [MCP server](/configuration/mcp) page.

## Access is scoped

Every token carries a role, from viewer through admin, and each tool requires a minimum role. Give an agent the **lowest role that lets it do its job**: a read-only agent that just inspects builds needs only `viewer`, while one that writes secrets needs `admin`. Tokens are scoped to a single organization, stored hashed, and revocable at any time. See [roles and security](/configuration/mcp#roles).