Runway MCP Puts Video Generation Inside Claude Agents
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, poll their status, and return.
PromptCrates Editorial
Staff Writer
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, poll their status, and return media URLs inside the same conversation, AlphaSignal reported on 25 September 2026, citing Runway’s MCP documentation. The remote endpoint at https://mcp.runwayml.com/mcp works with Claude, ChatGPT, Cursor, Replit, and Grok Bot via Streamable HTTP and OAuth, so teams authorize a Runway account instead of pasting API keys into client configs.
How the connector turns chat into a media loop
MCP gives agents a standard way to discover and call external tools. Runway’s hosted server translates those calls into generation jobs, handles asynchronous polling, and returns an output URL while saving the asset to the user’s Runway library for audit. In practice, a filmmaker or marketer can keep the brief, shot list, references, and revision notes in one thread while Claude decides when to call a generation tool.
As of September 2026, Runway’s connector page lists video models Gen-4.5, Gen-4 Turbo, Seedance 2.5, Kling 3.0, and Veo 3.1, plus image models Gen-4 Image, GPT Image 2, and Nano Banana Pro. The agent can ask Runway’s router to pick a compatible model or request one explicitly; availability still depends on the linked Runway plan. Generations bill against existing Runway credits, which makes spending controls a client-side permission problem as much as a creative one.
Setup paths are deliberately short. Claude.ai and Claude Desktop custom connectors accept the hosted URL after an OAuth browser flow. Claude Code can register the server with claude mcp add --transport http runway https://mcp.runwayml.com/mcp. Cursor users add a .cursor/mcp.json entry pointing at the same URL. That pattern matches how coding agents already attach browsers and repos, and it sits beside Runway’s earlier editor plugins covered in Runway DaVinci Resolve plugin and Runway Adobe Premiere After Effects plugin stories.
Why MCP matters more than another API wrapper
Creative teams have long chained Runway through custom scripts. A hosted MCP connector removes credential handling, job polling, and model routing for supported clients, so the agent loop can plan, generate, inspect, and revise without a separate microservice. AlphaSignal highlights production patterns that fit: reading a product page to draft a shot list, generating a hero video beside site copy in Cursor, turning one product still into several shots with shared references, and batching image variants for campaign review.
Those workflows still fail in familiar ways. Video jobs are asynchronous and can time out; models may be unavailable on a plan; moderation can reject prompts; outputs vary across retries; agents that browse untrusted pages can pick up hostile instructions. Runway’s library trail helps attribution, but teams that need per-user spend caps should confirm which identity metadata their plan records before letting an agent fire unsupervised batches. PromptCrates’ earlier Runway Dev MCP for coding agents coverage already flagged the developer-facing MCP surface; today’s story is the media-generation endpoint landing inside general Claude and ChatGPT sessions.
Editors comparing NLE plugins versus agent connectors should treat them as complementary. Premiere and Resolve plugins keep finishing inside the timeline; MCP keeps ideation and rough generation inside the chat where briefs live. Many shops will use Claude to explore shots, then pull approved URLs into Adobe Premiere generative media timeline workflows for final cut.
What production teams should configure on day one
Four configuration steps reduce surprise bills and orphaned jobs. First, require human confirmation before large batch generations or open-ended revision loops. Second, surface job identifiers in the chat so retries do not duplicate completed work. Third, set an explicit fallback model when the router’s first choice is plan-locked. Fourth, separate retrieved web content from tool instructions so a scraped product page cannot silently authorize paid renders.
Creative leads should also decide which models are allowed for client work. Gen-4.5 and Veo 3.1 may differ on motion fidelity and licensing posture; brand guidelines need a written map, not a hope that the router guesses correctly. Security reviewers should treat OAuth-linked Runway access like any other SaaS with generative spend: revoke unused clients, audit library shares, and log which agent identity triggered each job.
Agencies that already standardize on Claude for copy and Cursor for front-end work gain the most immediate leverage: the same agent that drafts landing-page copy can request matching motion tests without leaving the workspace, then hand approved URLs to an editor who finishes in Resolve or Premiere. That handoff only works if producers keep a shared naming convention between Runway library titles and shot-list IDs in the chat, otherwise revision threads lose track of which frame was client-approved.
Primary reporting for this article: AlphaSignal’s 25 September 2026 newsroom piece on Runway’s MCP media connector, cross-checked against Runway’s MCP documentation at runway.com/mcp. Facts stay anchored there: hosted endpoint; Claude Opus 5.5 and multi-client support; OAuth without API key paste; listed video and image models; library audit trail; Claude Code and Cursor setup commands; credit billing to the connected account.
