The After Effects MCP server, explained
An After Effects MCP server lets an AI client such as Claude, ChatGPT or Cursor read the comp you have open and build native, editable layers and keyframes in it. Prism is a hosted one: you connect by address.
| You bring | Your own AI client |
| Prism brings | The After Effects tools and the connection to your open comp |
| You get | Native layers and keyframes, not a script and not a rendered clip |
| You keep | Your subscription. Prism takes nothing for the model that drives it |
What is an After Effects MCP server?
MCP, the Model Context Protocol, is the open standard AI clients use to call outside tools. An After Effects MCP server provides those tools for Adobe After Effects: your AI client uses them to read the project you have open, then create and change its compositions, layers, keyframes, text and expressions. With Prism, the work splits three ways:
| Part | What it does |
|---|---|
| Your AI client: Claude, ChatGPT, Cursor | Where you ask |
| Prism | The MCP server, online at one address |
| After Effects | A panel runs each request in your open comp |
Is there an official After Effects MCP server?
No. Adobe connects several of its apps to AI clients, but not After Effects. Adobe for Claude, launched as Adobe for creativity, is Adobe's own MCP server in Claude and covers Acrobat, Photoshop, Lightroom, Illustrator, Firefly, Premiere, Adobe Express, InDesign and Stock. Adobe's AI integrations page, for Claude, ChatGPT, Gemini and Slack, does not mention After Effects either. Adobe's AI Assistant is built into After Effects, in public beta since 8 September 2026, and it is not a server your own AI client connects to.
As of October 2026, every After Effects MCP server, Prism included, comes from a third party.
Hosted or local: two kinds of After Effects MCP server
The kind decides what you install and which AI clients can reach it.
| Local server | Hosted server, such as Prism | |
|---|---|---|
| Where it runs | On your computer, beside After Effects | Online, at one address, with a panel in your After Effects |
| What you install | The server, often with Node.js and a build step, and for most a panel or script in After Effects | A panel or an installer, then one address in your AI client. Nothing to clone or build |
| claude.ai | Not directly. It connects from Anthropic's servers, which cannot reach your computer | Yes |
| ChatGPT on the web | Not directly, for the same reason | Through developer mode, with write actions on Business, Enterprise and Edu |
| Sign-in | Usually none | An account, approved once per client in your browser |
| Updates | You update it on your computer | New tools arrive from the server |
A local server wins on control: an open-source one lets you change its code and keeps running without any company's service. The servers themselves, named and dated, are in After Effects MCP servers compared.
What you need
| After Effects | 2024, 2025 or 2026, on macOS or Windows |
| A Prism account | Free: 100 MCP actions a month, no card. Pro: $6 a month or $60 a year, with unlimited MCP actions. See pricing |
| An AI client | Any MCP client that connects to remote servers and signs in through your browser. See connect your AI |
| Not needed | Node.js or an API key |
How do I install the After Effects MCP server?
With Prism, install one panel in After Effects, paste one address into your AI client, and approve the sign-in in your browser: about five minutes, once. The five steps are in Set up After Effects MCP. A local server is usually cloned and built, or run with npx, then added to your client's config.
Already pay for ChatGPT, Claude or Gemini?
Then you have the AI already. Prism reaches After Effects from the client you already use, so your assistant keeps what it knows about you: your past projects, your naming, the way you brief.
| If | What to know |
|---|---|
| You use ChatGPT | Use the desktop app or Codex, on Plus, Pro, Business, Enterprise or Edu. On the web it needs developer mode |
| You use Gemini | It connects through Antigravity. See connect Gemini |
| You have no AI plan yet | Claude's Free plan allows one custom connector, and Prism can be that one |
| You use a model through its API alone | It does not connect: the client holds the connection |
You choose the model, and you keep choosing
| Your client | ChatGPT, Claude, Codex, Cursor, Gemini through Antigravity, Kimi Code, Qwen Code, OpenCode, Cline, Windsurf, VS Code with Copilot or Zed |
| Any model | Prism cannot see which model drives your client, so the tools and your allowance are the same for every model, including one you host yourself |
| Any time | Add the same address to as many clients as you like and switch whenever you want |
A stronger model plans better, which shows up as fewer wasted requests, not different tools. Prism records which client connected, by name and version, as product analytics.
What you are left with
The output is your project, not a picture of it.
| Layers | Native. Select, move, restyle |
| Keyframes | Real, with eases you can drag in the graph editor |
| Text | Still text. Retype it, change the font |
| Expressions | In the property's expression field, readable and yours to edit |
Nothing depends on Prism afterwards: cancel your plan and the work stays exactly as it is.
The loop
| Ask | Describe the change in plain words, in the client you already use. You never see code unless you ask for it |
| Read | Before it changes anything, the agent reads which compositions exist and the layers and timing of the one you name |
| Build | Prism checks every request before it runs, and one that fails the check changes nothing. One ask can build a whole comp: every layer, laid out from the real size of text and shapes, animated with eases |
| Check | The agent can look at frames of what it built before it reports back, so it checks its own work instead of assuming it landed |
Common questions
It depends on the server, and local servers usually ask for none. With Prism, yes: you sign in through your browser, with Google or an email code, once in the panel on each computer and once in each AI client. There is no API key, and the address carries no secret.
Yes. Prism works in the project you have open, through its panel, which opens from Window ▸ Extensions ▸ After Effects MCP and has to stay open while you work. It does not work on a closed project.
Yes. The ChatGPT desktop app and Codex connect to Prism with one address, in about five minutes, once, on Plus, Pro, Business, Enterprise and Edu, the plans OpenAI lists for MCP. ChatGPT on the web needs developer mode, with write actions on Business, Enterprise and Edu. Steps: Connect ChatGPT to After Effects.
Yes. Claude desktop, Claude on the web and Claude Code connect to Prism, and adding it once in the Claude app also covers Claude Cowork and Claude Design. It takes about two minutes, once. Steps: Connect Claude to After Effects.
After Effects 2024, 2025 and 2026, on macOS and Windows. After Effects 2024 needs panel version 1.4 or later, and releases earlier than 2024 are not supported. See compatibility.
Yes. Open-source servers are free to install and run on your own computer, and After Effects MCP servers compared lists them. Prism Free costs $0 and needs no card: every After Effects tool, with 100 MCP actions a month. Prism AI is a Pro feature.
No. Prism is a hosted MCP server, and what you install is one signed panel. The tools run on Prism's server, so new capability arrives without reinstalling anything.
Yes. Every request lands as one step in After Effects history, so a single undo reverts it, and several requests become one step when you ask for them as one job. Checkpoints are not automatic: ask for one before a large change and Prism saves a timestamped copy of your project.