Set up the Runway MCP connector in Claude and ChatGPT, pick the right model, and generate Runway videos and images without leaving the chat. This article walks through each step, the credit rules, prompt formats that work, and fixes for the most common connection errors.
Ask Claude for a ten second shot of a coastal road at golden hour, and a finished clip lands in the same chat a few minutes later. That is what a working Runway MCP Claude and ChatGPT setup gives you: one conversation where the assistant writes the brief, calls Runway's models, and hands back the video. No tab switching, no export step, no pasting prompts between apps. The connection itself takes about five minutes. What takes longer is picking the right model, writing prompts that survive a render, and reading the error message when the connector goes quiet.
This article walks through the whole path. You will connect Runway to Claude and to ChatGPT, see which models are on offer, write a first prompt that works, and fix the usual failures. Near the end there is a hands-on section for running Gen 4.5 on Picasso IA, for the days you want that model without any connector at all.
💡 Quick answer: In Claude, add
https://mcp.runwayml.com/mcpas a custom connector. In ChatGPT, add the Runway app. Sign in with your Runway account either way, and every generation is billed to your existing Runway credits.
MCP stands for Model Context Protocol, an open standard that lets an AI assistant call tools outside its own chat window. Runway announced its MCP server on May 27, 2026, and it works with Claude, ChatGPT, Cursor, and other MCP compatible agents. Once connected, the assistant can generate images and videos on your Runway account and return them in the same thread.
The assistant does the thinking and Runway does the rendering. Claude or ChatGPT reads your product page, rewrites your rough idea as a proper shot description, picks a model, and submits the job. Runway renders it and saves the output to your Runway library. You never copy a prompt from one app to another, which is where most people lose twenty minutes per session.
The connector lives at https://mcp.runwayml.com/mcp. Authentication runs through your Runway account, so there is no API token to paste for Claude or ChatGPT. Developers billed through the Runway API use a separate hosted endpoint, https://dev.runwayml.com/mcp, so check which kind of billing you actually have before you paste a URL.
Two minutes of checking saves an hour of confusion later. Most failed setups come down to a plan limit or an unfinished sign in, not a broken server.
Runway states that MCP uses your Runway credits, just like the Runway app. Cost depends on the model, the resolution, and the settings you pick. Newer models are gated by plan tier, so a model you read about may not appear on a small plan. A long clip on a premium model can also burn through a large share of a small monthly credit pool, which is why heavy users end up on higher tiers.
On the assistant side, custom connectors in Claude generally need a paid plan, so confirm yours supports them before you start.
| Item | Why you need it | Where to check |
|---|---|---|
| Runway account with credits | Every render is billed here | Runway account settings |
| Paid Claude plan | Custom connectors are not on every plan | Claude plan page |
| ChatGPT account with apps enabled | Runway is added as an app | ChatGPT settings |
| Pop ups allowed in your browser | The sign in step opens a new window | Browser site settings |
The same steps work in the Claude web app and in Claude Desktop, since both read your connector list from your account.
Runway as the name and paste https://mcp.runwayml.com/mcp as the URL, then click Add.💡 Menu labels shift as Anthropic updates the app. If you cannot find Customize, look for a connectors or integrations entry under settings. The URL is what matters.
Start with a free sanity check instead of a render. Type: "List the Runway models available on my account." If Claude answers with a model list, the sign in worked and the tools are live. If it says it has no Runway tools, the sign in was not finished: disconnect, reconnect, and approve the access screen again.
Then run one short clip: "Using Runway, generate a 5 second 16:9 video of waves rolling onto an empty beach at sunrise. Tell me the model and duration before you start." The last sentence matters. It makes the assistant show its plan, which keeps you in control of credits.
ChatGPT treats Runway as an app, so there is no URL to paste in the common flow.
OpenAI renames and moves this menu often. Runway's own MCP page currently points to chatgpt.com/plugins, where you search for Runway, click Install Plugin, choose Sign in with Runway, and then mention @Runway in any chat. Both routes end in the same place: your Runway account linked to ChatGPT, with no API token involved.
💡 If ChatGPT answers a video request with plain text and no render, the Runway app was not selected for that chat. Mention it by name, or select it from the tools or apps menu if your version has one, then resend the request.
Runway's announcement lists its own Gen 4.5 alongside partner models. Availability depends on your plan tier, and different Runway pages name different Seedance versions, so treat any list in an article as a snapshot. The reliable method is to ask the assistant for the live list before you choose.
| Model | Type | Good first use |
|---|---|---|
| Gen 4.5 | Video | Cinematic motion from text or a start image |
| Gen 4 Turbo | Video | Fast image to video drafts |
| Seedance 2.0 | Video | Clips that include built in audio |
| Kling v3 | Video | Cinematic scenes |
| Veo 3.1 | Video | Realistic scenes with sound |
For stills, the list includes GPT Image 2 and Nano Banana Pro. Runway's open source server notes recommend Nano Banana Pro as the default image pick, and that is a sensible starting point when you do not care which model draws the frame. Pick GPT Image 2 when you need product mockups or a layout with readable words.
A useful habit: generate a still first, approve it, then ask for a video that starts from that still. You spend a few credits on the image and avoid paying for a full clip that starts from the wrong frame.
Connecting the tool is the easy half. The prompt decides whether the clip is usable or a coin flip.
Treat each request as a mini shot brief, not a marketing sentence. Include these eight pieces:
Here is a brief that works as written:
Use Runway Gen 4.5. A ceramic stovetop coffee maker on a wooden table,
steam rising as it brews. Slow push in, shallow depth of field, warm
morning window light from the left. 10 seconds, 16:9, natural colors.
No text, no logos, no hands in frame.
If you want help drafting briefs before you send them, a chat model such as Claude Sonnet 5 or GPT 5.6 Sol on Picasso IA can turn a loose idea into this eight part format first.
Video generation is asynchronous. The assistant submits a task, Runway renders it, and the result arrives minutes later, longer for 10 second clips on premium models. The most common beginner mistake is resubmitting because nothing appeared yet, which creates a duplicate task and spends credits twice.
Instead, ask: "Check the status of the Runway task you just started." Assistants that support task lookups will report whether the job is queued, running, or finished, and hand over the link when it is ready.
Runway's own examples show where chat generation pays off fastest:
Keep spending predictable with three rules. First, tell the assistant to state the model, duration, and resolution and wait for your approval before every render. Second, draft with a fast model, then rerun only the winning idea on a premium one. Third, test every new prompt at the shortest duration. A 5 second test costs far less than a 10 second miss.
Almost every failure falls into one of the rows below.
| Symptom | Likely cause | Fix |
|---|---|---|
| Connector added, no Runway tools appear | Sign in never finished | Disconnect, reconnect, approve the access screen |
| A 405 error in a developer client | Wrong transport type | Use Streamable HTTP, not SSE |
| A model is missing from the list | Your plan tier does not include it | Ask for the live list and pick a fallback |
| Video never shows up | Render still running, or the chat lost the task | Ask for the task status instead of resubmitting |
| Generation refused | Out of credits, or content moderation | Check your balance, then adjust the prompt |
Runway's usual content moderation still applies in chat, so a prompt that is blocked in the app will likely be blocked here too.
Developer tools use the same hosted server. Add the URL in the client's MCP settings and sign in with your Runway account. Many clients read a JSON file shaped like this, though the exact file name and fields vary, so check your client's docs:
{
"mcpServers": {
"runway": {
"url": "https://mcp.runwayml.com/mcp"
}
}
}
Community setup notes report that the server speaks Streamable HTTP only, and that clients forcing an SSE connection can get a 405 back. If your client has a transport option, set it explicitly.
Runway also published an open source server, runwayml/runway-api-mcp-server, which ran locally with a RUNWAYML_API_SECRET value taken from the developer portal. Its tools included runway_generateVideo, runway_generateImage, runway_upscaleVideo, runway_editVideo, runway_generateAudio, and runway_getTask. Its last push was on September 30, 2026, and the repository is now marked archived on GitHub.
For a new setup, use the hosted server instead. If you still run the archived one, remember that the README warns images from the Runway API live only 24 hours at the generated link, so download your files the same day.
If you would rather skip connectors, Gen 4.5 is available directly on Picasso IA. It generates video from a text prompt, or from an optional starting image, in 5 or 10 second clips.
| Setting | Options | Tip |
|---|---|---|
| Prompt | Free text | Name the camera move and the light |
| Duration | 5 or 10 seconds | Test at 5, finish at 10 |
| Aspect ratio | Six choices | 9:16 for stories, 21:9 for intros |
| Starting image | Optional | Use an approved still to avoid a wrong first frame |
| Seed | Integer | Reuse it to repeat a result |
Picasso IA also offers its own MCP connections from your account page, with four models at the time of writing: Picasso IA Image, Image Editor Pro, Picasso IA Video, and Seedance 2.5 Lite, with a limit of five concurrent predictions. Read the plan details on the site before you rely on them.
The fastest way to see the difference between a vague prompt and a sharp one is to run both. Take the coffee maker brief from earlier, open Gen 4.5 on Picasso IA, and generate it at 5 seconds. Then remove the camera and lighting lines and run it again. Put the two clips side by side and you will see exactly what each line of a shot brief buys you.
Need a first frame? Create one with Picasso IA Image, fix details with Image Editor Pro, and then animate it. Pick a subject you care about, write the brief, and create your own images and your first clip on Picasso IA today.
