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

# MCP

> The Cloud MCP server lets any AI client drive Cyberun end-to-end.

The [Model Context Protocol](https://modelcontextprotocol.io)
(MCP) is how external AI agents discover and call Cyberun's
runtime: list workflows, submit tasks, watch progress, fetch
results, cancel runs. Cyberun ships a **cloud MCP server** that
speaks the official Streamable HTTP transport and works with
every mainstream MCP client out of the box.

<CardGroup cols={2}>
  <Card title="Cloud MCP" icon="cloud" href="/mcp/cloud">
    Hosted at `core.cyberun.cloud/api/v1/mcp`. 15 tools that
    wrap the team's full runtime API. Authenticated with an
    integration key (`sk-`). Use from anywhere.
  </Card>

  <Card title="AI skill" icon="sparkles" href="/skill">
    One shell command (`npx skills add cyberun-cloud/skills`)
    teaches Claude Code, Cursor, Windsurf, Codex CLI, or any
    other skill-aware agent how to use Cyberun without you
    re-explaining.
  </Card>

  <Card title="API reference" icon="terminal" href="/api-reference/introduction">
    The HTTP surface that MCP wraps. Useful when you need an
    operation MCP doesn't expose, or when you're not in an
    MCP-aware environment.
  </Card>
</CardGroup>

## When to use MCP

| You want…                                      | Use                                     |
| ---------------------------------------------- | --------------------------------------- |
| An AI tool to drive workflows from any machine | [Cloud MCP](/mcp/cloud)                 |
| Long-running session, many small operations    | [Cloud MCP](/mcp/cloud)                 |
| Programmatic access without an MCP client      | [REST API](/api-reference/introduction) |

## Standards compliance

The server implements the official Model Context Protocol
specification. It speaks
**Streamable HTTP** as defined in MCP `2025-03-26` and later
(including the `MCP-Protocol-Version` header required by
`2025-06-18+`), uses Bearer-token authentication on
`Authorization: Bearer …`, surfaces tool errors as
`CallToolResult` with `isError: true`, and emits
`notifications/progress` for long-running tools.

Verified-compatible clients:

* **Claude Code** (CLI and IDE)
* **Claude Desktop**
* **Cursor**
* **Windsurf**
* **VS Code** 1.101 and later (native HTTP MCP)
* **Codex CLI**
* Any client that supports the legacy stdio path via
  [`mcp-remote`](https://www.npmjs.com/package/mcp-remote)

## What you can do over MCP

The common workflow operations are:

* **List workflows** the team has published.
* **Submit a workflow** with parameters; receive a `task_id`.
* **Stream task progress** until terminal state, with MCP
  `notifications/progress` events along the way.
* **Fetch the result** — output metadata plus signed download URLs.
* **Cancel a task** that hasn't finished yet.
* **List the team's agents** and see which are idle.

The Cloud MCP additionally exposes container-service inspection,
file presign for workflow inputs, and workflow lookup by slug.

For the complete tool inventory with input schemas, see
[Cloud MCP → Tools](/mcp/cloud#tools).

## Authentication, briefly

* **Cloud MCP**: `Authorization: Bearer sk-...` — an integration
  credential issued from Cyberun Cloud's access settings. The
  team scope is baked into the key; no `X-Team-ID` header is
  required for MCP requests.

Full auth details are on the server's page.

## Next

<CardGroup cols={2}>
  <Card title="Connect Cloud MCP" icon="cloud" href="/mcp/cloud">
    Endpoint, auth, every tool, per-client install snippets.
  </Card>

  <Card title="Install the AI skill" icon="sparkles" href="/skill">
    Teach the agent the API + MCP layout in one command.
  </Card>

  <Card title="MCP spec" icon="book" href="https://modelcontextprotocol.io">
    Background reading on the protocol itself.
  </Card>
</CardGroup>