Screel now runs inside ChatGPT and Claude

Screel has always been a place you go to: open the editor, write a script, shape your scenes, render. That's a good workflow when you already know what you want to make. It's a worse one when the idea is still forming in a conversation.

So we built the other direction. Screel is now an MCP server — the same account, the same projects, the same renders, reachable from any assistant that speaks the Model Context Protocol. Today that means ChatGPT and Claude.

What that actually looks like

You describe an idea in chat. The assistant does the scriptwriting — breaking your idea into scenes, drafting the narration, picking a visual direction — and then calls Screel to build it:

create_project        → a new project appears in your Screel account
add_scenes            → the whole script, in one call
generate_scene_image  → every scene gets its visual
animate_scene         → still frames become motion (optional)
generate_voiceover    → narration in the voice you picked
render_project        → a finished video on a permanent CDN URL

Every one of those is a real Screel operation against your real account. When the render finishes, the video is in your project list on the web app, exactly as if you'd made it in the editor. Nothing is sandboxed or throwaway.

It's not a text wrapper

The part we care most about: the results are visual, in the chat.

We shipped seven interactive widgets that render inline while you work — a storyboard that shows your scene grid and updates live as images land, a per-job progress card with an inline preview of what just generated, a scene card for working on one shot at a time, and a video player for the finished render. You can tap a scene to blow it up, and act on what you see without retyping context the assistant already has.

Generation is asynchronous, so a tool call never blocks waiting on a render. The call returns a job immediately and the widget tracks it from there.

Images from the conversation come along

This one turned out to matter more than we expected. If there's already an image in the chat — one you dragged in, or one the assistant made — you don't regenerate it in Screel. set_scene_image attaches the file you're looking at directly to a scene, and it costs nothing. The image you liked in the conversation is the image in the video.

Worth being straight about a difference here. ChatGPT can generate an image in the conversation for Screel to pick up. Claude deliberately has no image generation of its own, so there the handoff starts from a file you upload. Either way generate_scene_image covers the scenes you haven't got a picture for — the handoff is a shortcut, not the only route.

Connecting your account

Linking is a normal OAuth flow — no API keys to paste anywhere. Screel runs a dedicated authorization server, and the assistant registers itself, sends you through a login and consent screen, and gets a scoped token back. If you don't have a Screel account yet, you can create one inside that same flow and land back in the conversation without a detour.

Two scopes, and you see both before you agree:

• screel.read — view your projects and videos
• screel.generate — create content, which spends credits

Where it runs

Screel isn't a ChatGPT integration that happens to also work in Claude. It's one MCP server. Every client that speaks the protocol connects to the same account, the same projects, and the same set of tools.

The third column is the one people miss. Anything that takes a remote MCP server URL — Claude Code, Cursor, Le Chat, Perplexity — can point at https://api.screel.ai/mcp and get the full pipeline. The one thing you give up is the visual layer: widgets need a client that renders them, so elsewhere the same tool calls come back as text. The videos are identical.

What it deliberately can't do

We shipped 19 tools and none of them delete anything. No deleting projects, no deleting renders, no destructive edits an assistant could talk itself into. Scene edits are non-destructive, and image refinements keep the previous version so you can revert.

Generative tools also report what they cost — credits spent and credits remaining come back with the result, so the assistant is quoting your real numbers rather than guessing. Account settings and billing stay on the web app, where they belong.

Where this is going

The honest framing: this is the first release, and it's a complete one rather than a preview slice. The full pipeline works — script, images, motion, voiceover, music, render, upscale — from inside a conversation.

What we want to learn now is which half of the workflow people actually want in chat. Our guess is that ideation and first drafts belong in the conversation, and precision editing belongs in the editor. But that's a guess, and the whole point of shipping is to stop guessing.

If you try it, tell us what broke. We're listening.