Runway brings media generation to Claude and other agents through MCP

Runway has released a hosted Model Context Protocol connector that lets Claude Opus 5.5 and other tool-capable agents submit image and video jobs, monitor them, and return results inside a conversation. Users can develop a storyboard, generate shots, inspect outputs, and request revisions while the agent retains the brief and references in its working context.

The connector operates at the client’s tool layer. Claude decides when to call a generation tool, Runway executes the job, and the client displays the returned media URL as a link or preview. The same endpoint can work with ChatGPT, Cursor, Replit, Grok Bot, and other clients that support remote MCP servers, Streamable HTTP, and OAuth authorization.

From prompt to media URL

MCP gives agents a standard way to discover and call external tools. Runway’s hosted server translates those calls into generation jobs, handles asynchronous task polling, and returns the output to the client.

1. The client loads Runway’s available tools and model options.
2. The agent selects a tool and sends the prompt, settings, and supported reference media.
3. Runway submits the job and monitors its status.
4. The connector returns a media URL, while the generated asset appears in the user’s Runway library.

As of September 2026, Runway’s connector page lists the following routed models:

The agent can ask Runway’s router to choose a compatible model or request one explicitly. Availability depends on the Runway account plan and may change as the catalog evolves. Prompts organized around output constraints such as duration, aspect ratio, tone, and shot style remain portable across models.

Connect once and authorize with OAuth

Runway’s OAuth flow links the MCP client to a Runway account, so no Runway API key needs to be copied into client configuration. Generations consume that account’s credits, which makes connector access equivalent to spending access for permission and approval purposes.

Claude’s two setup paths

Claude.ai and Claude Desktop builds that expose custom connectors accept https://mcp.runwayml.com/mcp as the server URL. After adding it, complete the Runway OAuth flow in the browser.

Claude Desktop installations that rely on local MCP configuration can use the mcp-remote bridge. This path requires Node.js and npx:

{
  "mcpServers": {
    "runway": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.runwayml.com/mcp"
      ]
    }
  }
}

Claude Code in one command

claude mcp add --transport http runway https://mcp.runwayml.com/mcp

Cursor with a project file

Cursor can register the remote server in .cursor/mcp.json:

{
  "mcpServers": {
    "runway": {
      "url": "https://mcp.runwayml.com/mcp"
    }
  }
}

ChatGPT, Replit, Grok Bot, and other hosts use their own connector or MCP settings. Access may depend on the client plan, workspace policy, administrator approval, and support for OAuth-backed remote servers.

Production workflows that fit the connector

Agent access is useful when planning, generation, and revision already happen in the same workspace. The conversation can hold the creative brief, shot list, reference assets, code, and copy while Runway handles the media jobs.

• Have a browsing-capable agent read a product page, draft a shot list, and generate launch assets.
• Create a hero video alongside matching site copy and front-end code in Cursor or Claude.
• Turn a product image into several shots while reusing references and visual constraints.
• Iterate on dialogue-driven character ads without manually uploading the same files for each request.
• Generate batches of image variants from a shared specification for experiments or campaign review.

Reference images can improve continuity when the selected model and client support them, although identity, lighting, composition, and product details can still drift between generations. Multi-shot work therefore requires output review and selective regeneration.

Costs and failure modes to design for

MCP-generated assets use Runway’s existing credit system and appear in the central Runway library. That provides a shared asset history, while teams requiring per-user attribution should verify which prompt, identity, and activity metadata their Runway plan records.

Where MCP changes the integration work

The connector turns Runway into a callable media service within an agent loop that can plan, generate, inspect, and revise. It also reduces custom work around credential handling, job polling, file transfer, and model routing for supported clients.

Developers still need to manage tool permissions, spending limits, asynchronous failures, model availability, moderation responses, and nondeterministic output. For teams that already plan campaigns or build products in Claude or Cursor, the connector keeps more of that workflow in one conversation while preserving generated assets in Runway.