How to connect any AI model to AtomCut.
Use the model you choose. Keep the creative work inside AtomCut.
in your browser.
An illustrated guide to the studio. Created by CodeIdeal. Product details link to AtomCut’s official documentation.
Start with the connection you need.
AtomCut can use a supported hosted provider, a local model, or a compatible API endpoint. You can also control the editor from an external MCP agent. “Any model” means a model available through those supported routes; it does not mean every model can perform every task.
Choose where you want to work. If you want the conversation beside your canvas, use the built-in Agents panel. If you already work in an agent app and want it to coordinate files, scripts, and the editor, use the MCP route.
Connect a model inside AtomCut.
- Open Agents. In the studio, choose Connect a model, or open Connections from the menu.
- Choose a provider. Add its API key inside AtomCut and select an available model. Hosted options include Anthropic, OpenAI, Google, OpenRouter, and other supported providers.
- Press Test. A successful test checks a real round trip through the connection. Then make it the chat connection you want to use.
Use the available model list rather than copying a model name from an old tutorial. New models and provider permissions change independently of the editor.
Test the workflow with a tiny, visible edit.
A successful connection test tells you the endpoint answered. Your first creative request should check whether the model can work sensibly with the project.
Start with a short, specific brief:
Create a five-second title card. Use the text “Made with care.” Keep it centered with generous margins. Animate it in gently, then hold it still long enough to read. Show me a frame before you finish.
Inspect the layer list, the timing, and the rendered result. Can you find the text and edit it? Does the entrance finish before the reading hold begins? Can you undo the request cleanly? Those checks tell you more than asking for an entire launch video first.
Once that works, increase the scope a little at a time. Ask for a second scene, a different aspect ratio, or an audio cue. A smaller request is easier to review and easier to correct.
Chat and media jobs are separate choices.
The chat model plans the work and calls tools. Image, video, voice, music, and transcription jobs may use separate providers. A media-only connection does not become the conversation model.
In a connection’s routing options, choose which jobs it should handle. Test one small job before starting a large generation. Give an image request an image-capable route and a voice request a voice-capable route.
A useful first brief separates the work clearly: “Use my existing footage. Make the edit and title with editable layers. Ask before generating any new media.” That leaves less room for an assistant to spend time or provider credits on something you did not need.
Use a local model or compatible endpoint.
AtomCut supports local connections such as Ollama and LM Studio, as well as OpenAI-compatible endpoints. Configure the local server and browser access using the current connection documentation.
The model still needs to suit the job. A text-only model cannot inspect an image as a vision model would. A small model may handle a short instruction well but struggle to plan a long sequence of tool calls. Keep an initial test short and check the result.
For a company gateway, ask the person maintaining it for the endpoint and model identifier it exposes. Don’t infer API compatibility from a model’s name. Use the editor’s connection test, then the small title-card exercise above.
Connect an external agent with MCP.
MCP lets an agent app discover AtomCut’s commands and documentation. The official bridge connects that client to a running editor. Add the AtomCut MCP server in your client, then use Agents → Use your agent app → Link in AtomCut.
The command below is the documented Claude Code setup. Other clients have their own configuration; use the official instructions for yours.
claude mcp add -s user atomcut -- npx -y @atomcut/mcp@latestKeep the editor open and linked while the agent works. Ask it to inspect the current composition first, make a small change, and render a frame. The bridge makes the tools available; the connected model still decides which tools to call.
You ask a connected agent for a concrete change. The language model chooses the next tool call; MCP is the connection protocol, not the model.
When a connection does not behave as expected.
The connection test fails.
Read the returned error first. Check the endpoint, model identifier, API key, and provider access. Change one thing at a time and test again so you know which setting mattered.
Chat works, but a media request does not.
Check the connection assigned to that type of job. Confirm the chosen provider and model support the requested operation. Try a small request before retrying a full scene.
The agent makes a messy edit.
Reduce the request to one scene or one change. Give a size, duration, exact text, and desired motion. Ask for a render and explain the specific problem you see before adding more work.
An external agent cannot find the editor.
Check that AtomCut is open and linked to the agent app, and that the MCP server is configured in the client you are actually using. Follow the bridge’s returned pairing instructions.
The result moves too much.
Ask for a reading hold, a restrained entrance, and fewer animated elements. Tell the agent which moment should carry the emphasis. A precise revision is more useful than “make it better.”
Build a brief the model can follow.
For the next request, write down the output size, duration, exact on-screen text, available footage, and one reference for the motion. Say whether the assistant should generate media or work only with the assets you already have.
Ask it to leave text editable, use clear layer names, and show frames at useful moments. Review the beginning, the main change, and the ending. Then adjust the few things that need attention.
AI agents are a Pro capability. Provider API usage is separate from an AtomCut license; a chat subscription alone is not necessarily an API credential. Check the current plan and provider terms before a large session.
Connect a model, then make one thing.
Start with a small composition you can inspect and edit.
Open AtomCut ↗Read the complete AtomCut studio guide →Written by CodeIdeal, the team building AtomCut. Setup references reviewed October 26, 2026. This article never asks for or stores your API keys; configure connections in the editor.