carvercat
Docs menu · Agent guide

Connect, inspect, edit and hand back.

A practical local MCP workflow built on reading current state, not on assumptions.

  1. Separate the app from the AI provider.
  2. Check the connection with a read.
  3. Watch the agent indicator.
  4. Let the agent discover what the app offers.
  5. Use the units from the web.
  6. Give the agent the Carvercat skill.
  7. Keep the edit bounded.
  8. Refresh after manual changes.
  9. Recover from disconnects and uncertainty.

Separate the app from the AI provider.

carvercat’s local MCP is free, like everything else in the app. A compatible AI agent can inspect and edit the same project you work on visually. Your AI client may have its own account, network and costs, and its own data handling.

Compatible AI clients can connect. Results depend on the client and model you use.

Check the connection with a read.

  1. Open carvercat and choose a practice project.
  2. If your client is listed under Connect to AI, use that. Otherwise follow the connection steps the app shows.
  3. Keep your other client settings. Restart the client if its setup asks you to.
  4. Ask for open comps and current state without editing.
  5. Try one simple edit, check it, Undo it, and confirm a saved and reopened test copy.

Editing the open project and headless batch work are separate modes. A connected agent does not mean a project is open and ready.

Watch the agent indicator.

The status bar shows a small robot: ready, working or paused. While an agent edits, the canvas glows. Each agent call is one labelled undo step.

Press Stop to pause the agent, and Resume to let it continue. Only you can pause and resume. The agent can read its state and recent actions.

Carvercat now has 338 MCP commands, 47 of them new in 0.2.0. They cover keys, layers, links, clips, tracks, views, transitions, selection and playback.

Let the agent discover what the app offers.

Ask the agent to list the available tools and resources, and to identify project, comp, object and range by their stable IDs rather than by position.

If the app refuses or skips an action, that is useful context. The agent should adjust its plan rather than work around the restriction.

Use the units from the web.

Parameters use units from HTML and CSS. A bare number is pixels (or degrees, or a factor), and a string carries its unit with no space, such as "24px" or "50%".

  • Size and position: pixels, percent or points, measured from the top-left of the frame, y down.
  • Colour: hex, rgb(), hsl() or a CSS name, as it appears on screen.
  • Time: frames, or text such as "1.5s" or "500ms".
  • Easing: names and curves from CSS, such as ease-out or cubic-bezier().
  • A value Carvercat cannot take is refused with a reason, what would be accepted and an example, and nothing changes.

Give the agent the Carvercat skill.

The free Carvercat skill teaches an agent motion design and careful editing: timing and easing, typography, transitions, sound, vertical versions, a verified library of presets and templates, and whole-job workflows. It is optional; the app describes its own tools without it.

  1. Download or clone github.com/carvercat/carvercat-skills.
  2. For Claude, copy its ai folder into your Claude skills folder and name it carvercat-motion. The repository’s README has a one-line installer.
  3. For other agents, point the agent at ai/SKILL.md. It reads only the references a job needs.
  4. Ask for the motion you want in plain words, then check the result in carvercat.

Keep the edit bounded.

  1. Describe the change and what must stay the same.
  2. Read state and anything the change depends on.
  3. Make the change with revision checks.
  4. Check the state and a visible or rendered result.
  5. On conflict, reread and replan.
  6. Stop at the agreed point and describe what changed.

Correct values do not guarantee it looks right. Delivery needs its own review of target, range and method.

Refresh after manual changes.

You can edit the same project by hand at any time. Before the next agent action, it should read current state again and keep your changes. Old cached wording should never overwrite a newer manual edit.

Recover from disconnects and uncertainty.

If no project is ready, open it in carvercat and retry a state read. After a timeout, check state, history and output before repeating an edit. If sending to an editor, check Resolve or After Effects first. A lost reply does not mean the action failed.

carvercat processes your project locally, but your AI client may send data to its provider. Share only what the task needs, review your client’s permissions and remove customer data from reports.