MCP Servers for Video Production: What Each One Actually Does (2026)

MCP servers for video production now cover most of the chain, one stage at a time. ScreenWeaver handles the script, the breakdown and the shot list. Blender MCP drives 3D previs. Runway and Higgsfield generate shots. A community server controls DaVinci Resolve for the edit. ElevenLabs' hosted server does speech. No single server takes you from page to final cut.

That last sentence matters more than the list. If you are new to the protocol, read what an MCP server is first, then come back. This page is the map: which server sits at which stage, who maintains it, how it signs you in, and what it can spend or break.

Disclosure: we build ScreenWeaver, and Frank Houbre, one of our co-founders, uses it on his animated series Lost Garden. Everything below about the other servers comes from their official pages, read and captured on 26 October 2026. Where we could not verify something, we say so.

How we picked and checked these servers

Search "video MCP server" and you get directories listing hundreds of entries, many of them weekend projects. We kept a server only if it met three conditions:

It does a job a film or video production actually has (write, plan, previs, generate, edit, sound).
It has an official page or a maintained repository we could read on the day.
We could state plainly how it connects and what it costs, or say that the page does not tell you.

That rules out a lot. Adobe, for instance: on 26 October 2026 we found no Premiere Pro MCP server published by Adobe, only several community projects with very different tool counts. We left them out of the table because we could not check their claims.

Map of a video production chain in seven stages, write, break down, storyboard prep, previs, generate, edit and sound, with the MCP server that covers each stage: ScreenWeaver for the first three, Blender MCP for previs, Runway MCP and Higgsfield MCP for generation, the community DaVinci Resolve MCP for the edit and the ElevenLabs hosted MCP for speech, each tagged official or community and remote or local

The production chain and the server that covers each stage, as of 26 October 2026. Diagram by ScreenWeaver.

Write and break down: ScreenWeaver

This is the stage most video MCP lists skip, because the generators get the attention. But every shot a generator makes starts as a line in a script and a row in a shot list, and that is where a lot of AI films fall apart.

ScreenWeaver's server lives at https://app.screenweaver.ai/api/mcp. You sign in with OAuth 2.1, no API key, and on the consent page you choose to share all your projects or only some. It exposes 25 tools: eleven that read and fourteen that write. Among them:

extract_file_text reads a PDF. The text layer is free; OCR only runs on scanned pages.
import_script loads a screenplay. With dryRun: true it shows the scenes, places and character links it would create, and writes nothing.
create_entities builds the story elements (characters, places, objects), and create_shots writes the shot list with framing, movement and duration.

The real example we keep coming back to is Dracula. Frank gave the assistant the whole novel as a PDF. What came back in the project: 11 characters, 16 places and 8 objects, each with its sheet, and 32 scenes broken down into 177 shots.

ScreenWeaver Storyboard tab on the Dracula project: story element sheets for Jonathan, Dracula, the coachman and the innkeeper on the left, shot S1 with its framing, crane down movement and action text in the centre, the screenplay on the right, and the shot timeline across scenes 1 to 3 at the bottom

The Dracula project in ScreenWeaver after the breakdown: story elements, shot S1 and the script side by side. Screenshot from screenweaver.ai/start.

Two limits. The server is for the Auteur and Studio plans only. And it generates nothing: no image, no video. The assistant prepares shots, prompts and references; you launch the generation yourself in the Storyboard or Production tab. Connection guides: Claude, ChatGPT, and Claude Code, Cursor and VS Code.

Storyboard prep: comic pages, sketches and references

Same server, different tools. If your source is visual (a comic page, a webtoon episode, a page of pencil thumbnails), import_media brings it in as PDF, PNG, JPEG or WebP, up to 25 MB. Then you have two ways to cut it into panels. With crops, the agent gives the boxes itself and nothing is charged. With findPanels, a vision model detects the panels for about 1 credit a page, and ScreenWeaver's own documentation cites a 20-page board at 6 panels a page that cost $0.097.

After that, add_reference_images attaches up to 8 images to a shot's panel, a story element or the art direction, and set_prompt rewrites the prompt the panel generator will use. The board images themselves are still generated in the app.

Previs and 3D: Blender MCP

Blender MCP is a community project, and its README says plainly that it is not affiliated with the Blender Foundation. It has two parts: an addon that runs a small socket server inside Blender (by default on localhost:9876) and an MCP server that relays your assistant's requests to it. It works with Claude Desktop, Claude Code, Cursor and Codex, among others.

For a filmmaker, the useful part is blocking a set or a camera move in 3D before generating anything. The tool that does the heavy lifting, execute_blender_code, runs Python written by the AI inside Blender. The README calls that "potentially dangerous" and tells you to save your work first. It also warns that the addon's socket has no authentication, so keep it on your own machine. An optional safe mode screens scripts for risky operations before they run.

Generate: Runway MCP and Higgsfield MCP

This is the crowded stage. Two official servers stood out on the day.

Runway MCP. Runway's page lists ChatGPT, Claude, Grok Bot, Cursor and Replit as clients. You sign in with your Runway account, no API key, and generations draw on your Runway credits exactly as they would in the app, at a cost that depends on model and resolution. The models named on the page include Gen-4.5, Seedance 2.5, Kling 3.0 and Veo 3.1, with availability tied to your plan. Results land in your Runway library.

Runway MCP landing page with the headline Runway MCP, the line Generate high-quality images and videos right from where you're already working, buttons to add Runway to ChatGPT and start with examples, and a Connect Runway in seconds section listing ChatGPT, Claude, Grok Bot, Muse, Cursor and Replit

Runway's MCP page, captured 26 October 2026.

One trap if you search GitHub: Runway's older API server (runwayml/runway-api-mcp-server), which ran locally with an API key, was archived on 30 September 2026. The repository now points developers to Runway's agents docs and everyone else to the hosted MCP.

Higgsfield MCP. Higgsfield runs its server at https://mcp.higgsfield.ai. You add it as a custom connector in Claude, connect and sign in. Its own guide, dated 8 May 2026, says the connector reaches more than 30 video and image models, with output up to 4K and clips up to 15 seconds, and that it uses your existing Higgsfield credits. Those figures are Higgsfield's; we did not benchmark them.

Higgsfield Creator Hub article titled How To Generate AI Videos Straight From Claude with Higgsfield's MCP, dated May 8, 2026, with an intro paragraph saying the connector exposes more than 30 video and image models, output up to 4K and clips up to 15 seconds

Higgsfield's MCP guide, captured 26 October 2026.

If you already pay for Higgsfield, there is a second route: run ScreenWeaver's video nodes on your Higgsfield account, covered in using Higgsfield credits in ScreenWeaver.

A generator server is good at one clip. It knows nothing about scene 12, the character's jacket in scene 3, or which shot comes next. That context has to come from somewhere, and it is the reason the planning stage above exists.

Edit: the DaVinci Resolve MCP server

There is no official Blackmagic MCP server that we could find. The most starred one we found is davinci-resolve-mcp by Samuel Gursky: MIT licence, 3.8k stars and 52 contributors on the day we looked. It drives Resolve through the official scripting API. By default it runs as a local process started by your MCP client, with no network listener.

README of the davinci-resolve-mcp repository on GitHub: the title DaVinci Resolve MCP Server, badges for version 4.15.3, 37 MCP tools with 390 in the full set, 18 advanced offline tools, DaVinci Resolve 18.5 and later and an MIT licence, a description of control over editing, media pool, render, grading, Fusion and Fairlight, and a screenshot of its local control panel

The davinci-resolve-mcp README, captured 26 October 2026.

It ships in two shapes: a compact server of 37 tools that groups operations behind parameters, and a granular one with 390 tools, one per API method. The areas covered include timeline editing, media pool and ingest, markers, grading, Fusion, Fairlight and render queueing.

Check your edition first. External scripting is a Resolve Studio feature. The README describes a bridge that reaches the free edition from inside the app on Resolve 21.0.x, and says its behaviour on 21.1 is unconfirmed. If you edit on the free version, test before you plan around it.

Try it free

Try Screenweaver for free on your script

It is free. Import your existing project, get a clearer view of your outline, and regain control of your story structure in minutes.

Start Free

Sound: the ElevenLabs hosted MCP

ElevenLabs moved to a hosted server at https://api.elevenlabs.io/v1/mcp, with OAuth sign-in and nothing to install. Read the page before you connect it for a film, though. It is built mostly for managing ElevenLabs voice agents (creating them, editing them, reviewing conversations). On the audio side, the only capability it lists is text-to-speech, returned as a short-lived download link.

ElevenLabs documentation page Hosted MCP server, subtitle Create and manage your agents from Claude and other MCP clients, with an overview paragraph saying the server exposes agent management tools, the server address https://api.elevenlabs.io/v1/mcp and a note that authentication uses OAuth with no API keys copied into the client

ElevenLabs' hosted MCP documentation, captured 26 October 2026.

The older local server (elevenlabs/elevenlabs-mcp) had a wider audio toolset, including transcription, audio isolation and speech-to-speech, and needed an API key. Its repository is archived and marked as no longer maintained. For the rest of a soundtrack (music, effects, the mix), we have a separate guide on sound for AI films.

The servers side by side

Server	Stage	Made by	Runs	Sign-in	What it spends
ScreenWeaver	Write, break down, storyboard prep	ScreenWeaver (official)	Remote	OAuth, pick projects	OCR pages and `findPanels` only; generation happens in the app
Blender MCP	Previs, 3D	Community	Local (addon + server)	None on the socket	Nothing by itself; some asset generators are paid
Runway MCP	Generate images and video	Runway (official)	Remote	Runway account	Runway credits, same as the app
Higgsfield MCP	Generate images and video	Higgsfield (official)	Remote	Higgsfield account	Higgsfield credits
davinci-resolve-mcp	Edit, grade, render	Community	Local	None (local process)	Nothing; needs Resolve Studio for external scripting
ElevenLabs hosted MCP	Speech, voice agents	ElevenLabs (official)	Remote	OAuth	Not stated on the page

Comparison table of six MCP servers for video production checked on 26 October 2026: ScreenWeaver, Blender MCP, Runway MCP, Higgsfield MCP, davinci-resolve-mcp and the ElevenLabs hosted MCP, with columns for stage, maker, official or community, local or remote, sign-in method, what it can spend and what it can change on your side

The same six servers with the risk column added. Diagram by ScreenWeaver, facts as of 26 October 2026.

Before you connect one: four questions

Most of the risk in this list does not show up in feature tables. It sits in what a server is allowed to do once your assistant has it. Four questions sort them quickly.

Does it run code on my machine? Blender MCP does, by design. A local server is not unsafe in itself, but you are trusting whoever wrote it.
Can it spend money? Runway and Higgsfield spend credits every time a clip renders, and an agent can fire several in a row. ScreenWeaver asks for confirmation before paid steps: an OCR job above 20 pages, or panel detection above 20 pages.
Can it delete? In ScreenWeaver, delete_shots returns a preview and waits for confirmation, and import_script keeps the previous draft in History. Check the equivalent before you point any server at a real project.
Who maintains it? Runway and ElevenLabs both archived their first MCP server between August and September 2026. A community repository can stop overnight. Look at the last release date before you build a workflow on it.

Decision diagram with four questions to ask before connecting an MCP server to a film project: does it run code on my machine, can it spend credits, can it delete or overwrite, and who maintains it, with the servers sorted under each answer and the safeguard to use, such as safe mode, one test clip before a batch, a preview before deletion or a pinned release

Four questions before connecting a server to a real project. Diagram by ScreenWeaver.

Which ones you actually need

Fewer than you think. For an AI short, the combination that makes sense today is one planning server and one generator: ScreenWeaver for script, story elements and shots, then Runway or Higgsfield if you prefer generating from the chat. Add the Resolve server only if you already cut in Resolve Studio, and Blender MCP only if your film has sets worth blocking in 3D.

What we would not do is connect six servers to one assistant and ask it to make a film. Each handoff is still a file or a list, and each one deserves your eyes. To start with the planning end, the ScreenWeaver MCP page has the server address and the install buttons.

FAQ

Is there an MCP server that makes a whole film from a prompt?

Not that we found on 26 October 2026. Each server covers a stage. Planning servers like ScreenWeaver structure the script and shots, generators like Runway or Higgsfield render clips, and an editing server like the DaVinci Resolve one assembles them. Moving work between them is still on you.

Do MCP servers for video cost money?

The protocol is free. What costs money is the service behind the server. Runway and Higgsfield draw on your credits for every generation. ScreenWeaver's server requires an Auteur or Studio plan, and only OCR on scanned pages and automatic panel detection consume credits. The community Blender and Resolve servers are free, though Resolve's external scripting needs Resolve Studio.

Can ScreenWeaver's MCP server generate video?

No. It writes to the script, story elements, shots, prompts and references. You launch image and video generation yourself in the Storyboard or Production tab, where you see the cost before it runs.

Which MCP server works with DaVinci Resolve?

The most starred one we found is the community project davinci-resolve-mcp by Samuel Gursky, under MIT licence. It controls Resolve through the official scripting API, which needs Resolve Studio. Its README describes a bridge for the free edition on version 21.0.x and calls its behaviour on 21.1 unconfirmed.

Is there an official Premiere Pro MCP server?

We found none published by Adobe on 26 October 2026, only community projects. Check Adobe's developer documentation before relying on one, and treat third-party tool counts as unverified.