ChatGPT Runway MCP: Make AI Videos in Chat (2026)

ChatGPT cannot render video on its own, but it can now drive Runway from inside a chat. Connect Runway once, describe the shot you want, and the finished clip comes back in the same conversation — no API key, no tab-switching, and no separate prompt for every model. This guide walks through that whole loop: connecting Runway to ChatGPT, writing shot prompts that cut together, picking a model for graphics and motion, budgeting credits, and assembling the clips into a finished video.

Quick answer: Connect Runway to ChatGPT, ask @Runway for a shot, and turn the clips into a finished video. Setup, prompts, models, costs and assembly — no API key needed.

What is the ChatGPT + Runway MCP workflow?

MCP — the Model Context Protocol — is the open standard that lets a chatbot call tools on another service. Runway publishes a hosted MCP connection, so ChatGPT can call Runway’s generation tools the way it calls any other plugin. You stay in chat; Runway does the rendering; your Runway account pays the credits. The important mental model is that ChatGPT is the director and Runway is the camera crew: ChatGPT plans the shots and writes the prompts, but the footage itself is made — and billed — by Runway.

That division explains the workflow’s limits too. ChatGPT hands you short clips, not a finished, narrated film. Voiceover, captions, music and the final cut are still an editing step after generation, covered in step 6 below. And because every generation spends Runway credits, the cheapest habit in this workflow is also the most important one: approve a shot list before you generate anything.

What do you need before you start?

Three things, and nothing to install on your computer for the ChatGPT route:

A ChatGPT account with access to plugins/connectors. If your account is managed by a workplace, an admin may need to approve the Runway connection first — Runway’s own help page says exactly that for company and enterprise accounts.
A Runway account with credits. Generation through the connection uses the same credits as the Runway app, and Runway’s help page states that its Explore Mode free generations are not supported through MCP — only credit-paid generations work here.
One decision made in advance: the video’s shape. Pick the aspect ratio (16:9 landscape or 9:16 vertical), the total length (30 seconds is a good first target), and the number of shots (five or six shots of about five seconds each). Every later prompt should repeat those three facts, or the clips will drift apart in framing and style.

How do you connect Runway to ChatGPT?

Follow Runway’s official ChatGPT instructions, verified on 3 November 2026:

Go to ChatGPT’s plugin directory and search for Runway. You can also open the Runway plugin page directly from Runway’s help guide.
Select Install Plugin, then Sign in with Runway to connect your Runway account. There is no API key to copy — sign-in is handled by Runway’s OAuth flow.
In any chat, mention @Runway and describe the image or video you want. Attach a reference image if you have one — image-to-video keeps a character, product or location far more consistent than text alone.

If you later use a different MCP client (Claude, Cursor or Replit), the same Runway server is reached at https://mcp.runwayml.com/mcp. Runway’s help page notes that this server supports Streamable HTTP only, not SSE or stdio, so if a client offers a transport choice, pick HTTP. That URL is the generic connector address; inside ChatGPT itself, use the plugin install above rather than pasting the URL.

How do you get ChatGPT to write a shot list that actually cuts together?

The mistake most first attempts make is asking for “a video about X” in one sentence. Runway returns one short clip per call, so ChatGPT needs a shot list, not a theme. Ask for the list first, approve it, and only then ask for generation — that two-step order is what stops credit burn on shots you never wanted.

Start with a brief that locks the look for every shot:

Plan a 30-second vertical video about a small bakery at dawn.
Give me 6 shots of 5 seconds each, 9:16.
Keep one look across all shots: warm morning light, handheld camera,
no text on screen, no voiceover in the clips.
For each shot, write: shot number, one sentence of action,
camera movement, and the exact prompt to send to @Runway.

When the list comes back, check four things before generating: does each shot move the story forward, does the camera direction change just enough to feel edited (not a jump cut every time), is the same character or product described identically in every prompt, and is there any on-screen text the model will garble? Fix the list in chat first — editing text is free, regenerating video is not.

Then generate one shot at a time, naming the model when it matters:

@Runway, generate shot 2 with Gen-4.5, image-to-video,
using the bakery photo I attached: the baker slides a tray
of croissants into the oven, slow push-in, warm morning light,
5 seconds, 9:16, no text on screen.

If a clip comes back wrong, change one variable only — the action, the camera move, or the model — and say so explicitly (“same shot, but slower camera, no zoom”). Changing three things at once makes it impossible to learn what fixed it, and it burns credits on guesses.

Which model and settings should you pick for graphics and motion?

Runway’s own help page says the agent can use models including Seedance 2.0, Kling, Gen-4.5, Veo and its image models, and that it can pick a model for you or use the one you name — with availability depending on your Runway plan. An independent comparison of official video MCP servers (Revid, checked 22 October 2026) lists the same Runway roster as Gen-4.5, Veo 3.1, Seedance and Kling. Treat the roster as plan-dependent, not guaranteed: check your own account before promising a client a specific model.

Model (via Runway)	Best for	Watch out for
Gen-4.5	Your default for realistic, cinematic shots and consistent characters from a reference image	The most credit-hungry of Runway’s own models — budget it first (see costs below)
Gen-4 Turbo	Fast drafts and throwaway tests before you spend on a keeper take	Softer detail than Gen-4.5; use it to test the idea, not to finish the video
Seedance	Fluid, expressive motion where movement matters more than photorealism	Can overshoot on motion — ask for “slow, steady” movement explicitly
Kling / Veo (routed through Runway)	Matching a look you already know from those models, without leaving ChatGPT	Availability and cost vary by plan; confirm both before building a series on one

For motion quality, three settings habits matter more than model choice. First, describe one action per shot — a model asked to do three things in five seconds does all of them badly. Second, name the camera move (“slow push-in”, “static tripod”, “gentle pan left”) instead of leaving it to chance; unnamed camera moves are where the “AI wobble” comes from. Third, keep motion verbs small for product and face shots (“steam rises”, “she glances up”) and save big moves (drones, spins, chases) for wide establishing shots where a wobble reads as energy, not error.

How much will a 30-second video cost?

Runway bills in credits, and the number that decides everything is Gen-4.5’s rate: 12 credits per second, from Runway’s published pricing as re-checked by independent pricing roundups in September–October 2026. That means a 5-second Gen-4.5 clip costs 60 credits and a 10-second clip costs 120 — before a single retry.

Runway plan	Monthly price	Credits per month	Gen-4.5 seconds that buys
Free	$0	125, one time only	About 10 seconds — and free-plan videos carry a watermark
Standard	$15 ($12 billed yearly)	625	About 52 seconds
Pro	$35 ($28 billed yearly)	2,250	About 187 seconds
Max	$95 ($76 billed yearly)	9,500	About 791 seconds

Work the budget backwards from the finished video. A 30-second video at one take per shot needs 30 seconds of footage: 360 Gen-4.5 credits. Plan for two takes on half your shots and the honest budget is about 450–540 credits — roughly a Standard month’s allowance for one video, or a comfortable fraction of a Pro month. Two caveats, both labelled: first, these plan figures come from Runway’s pricing page as summarised by third-party roundups checked in September–October 2026, not from a live checkout on 3 November, so re-check before subscribing. Second, several of those roundups list MCP access from the Pro plan upward, while Runway’s own help page does not state a plan gate for the connection — treat Pro as the safe assumption, and confirm in your account if you are on Standard.

How do you turn clips into a finished video?

Runway hands ChatGPT clips; the finished video is still yours to assemble. The assembly below works in any editor (CapCut, Premiere Pro, DaVinci Resolve) or, for a scripted route, an editing API such as JSON2Video, which lets a chatbot place each clip in a video element and add voice and subtitles around it:

Download every keeper immediately. Generator clip links can expire or require a login later, so move each approved clip into your own folder or media library before you edit — this is also JSON2Video’s own advice for assembly workflows.
Order the clips to the approved shot list and trim heads and tails so each shot lands on its story beat. Cut on movement, not on silence: a shot that ends mid-gesture feels longer than it is.
Add one voiceover track and one music bed. Record the voiceover to the final cut (not before it), keep music low enough that the voice stays forward, and mute the clips’ own generated audio unless it is a deliberate sound effect.
Add captions last, from the voiceover script — not from auto-transcription of the AI audio. Burned-in captions should sit clear of faces and of any product detail the viewer needs to read.
Export once, at the ratio you chose at the start: 1080p for landscape or vertical delivery, with the same frame rate across every clip. Upscaling a mixed-ratio timeline is the fastest way to make a clean video look patched together.

What goes wrong, and how do you fix it?

Characters and products change between shots. This is the most common failure, and the fix is image-to-video from a single reference frame rather than re-describing the subject in text every time. Keep one hero image and reuse it as the starting frame for every shot that features that subject.

Motion looks frantic or warped. Shorten the action to one verb and slow the camera down in the prompt. If a shot still warps, switch to a calmer model (Gen-4 Turbo for the test, then Gen-4.5 for the keeper) instead of re-rolling the same prompt and hoping.

Text on screen comes out garbled. Do not ask a video model to render words. Generate clean footage with “no text on screen” in every prompt, then add titles and labels in the editor, where text is actually text.

Credits vanish on retries. Every generation bills, including the takes you discard — that is the single most expensive misunderstanding in this workflow. The approval gate (shot list first, generation second, one variable changed per retry) is not bureaucracy; it is the budget.

The connection itself fails. Runway’s help page gives three quick checks: confirm the transport is Streamable HTTP (SSE returns a 405 error), re-authenticate from the client’s MCP tools page if sign-in has expired, and, on a managed workspace, ask an admin whether the connector is allowed at all.

Is this workflow worth it, or should you just open Runway?

Open Runway directly if you are making one polished clip and you already know the prompt — the app gives you every slider, and chat adds nothing. Use the ChatGPT route when the work is conversational: planning a series, iterating a shot list, reusing one brief across many videos, or handing a non-editor a workflow they can run in plain language. Its real advantage is not speed per clip — generation takes the same time either way — but that the brief, the shot list, the prompts and the revision history all live in one chat thread you can reopen next week and reuse. For a one-off clip, that thread is overhead. For a weekly series, it is the whole point.

One honest limit to carry into that decision: this article is desk-researched from Runway’s official help page, an independent MCP-server comparison and pricing roundups, verified on 3 November 2026 — it is not a hands-on test in a paid Runway account. Model availability, the MCP plan gate and credit rates can all move with a pricing-page update, so check your own account’s current plan and credit balance before you commit a client’s deadline to this workflow.

Frequently asked questions

Can ChatGPT generate video by itself?

No. ChatGPT has no video-generation model of its own. Through the Runway connection it writes prompts and calls Runway’s tools, and Runway’s models make and bill the footage.

Do I need an API key to connect Runway to ChatGPT?

No. In ChatGPT you install the Runway plugin and sign in with your Runway account; Runway’s help page states explicitly that no API key is required. The generic MCP URL is only needed for other clients such as Claude, Cursor or Replit.

Which Runway plan do I need for MCP?

Runway’s own help page does not state a plan gate, but third-party pricing roundups checked in September–October 2026 list MCP from the Pro plan ($35 a month) upward. Treat Pro as the safe assumption and confirm in your account before subscribing for this feature alone.

How long are the clips ChatGPT can make through Runway?

Short — a few seconds per generation, sized to the model you choose (Gen-4.5 clips in the app run a few seconds each). A finished video is always several clips stitched together, which is why the shot list comes before generation.

Can I use the free Runway plan with ChatGPT?

Only in a limited way: the free plan gives 125 one-time credits, its videos carry a watermark, and Explore Mode free generations are not supported through MCP. For anything you intend to publish, a paid plan is the realistic starting point.

Sources and methodology

Method: on 3 November 2026, this guide studied Runway’s official MCP help page, an independent comparison of 18 official video MCP servers, a Claude-video pricing and assembly guide, and a recent YouTube Short demonstrating the ChatGPT + Runway connection, then cross-checked plan and credit figures against pricing roundups that re-read Runway’s own pricing page in September–October 2026. Where a figure comes from a secondary roundup rather than Runway’s own page, the article says so. No creator screenshot or clip is reproduced in this article, so no creator visual requires a caption credit here; the Short below is linked, not re-uploaded. Instagram was covered through public search only — the Instagram connector remains disconnected at the user’s direction — and no named Instagram guide for this exact workflow was found in the 24–72-hour window, so this article leans on the YouTube demonstration and the official docs instead.

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.