Learn / Development guide

Connect, inspect, edit and hand back.

A practical local MCP workflow that uses current state and discovery rather than assumptions about the installed app.

Development guide · source-reviewed foundation · public-package walkthrough pending

Separate the app from the AI provider.

carvercat’s local agent integration is included at no charge. A compatible client can inspect and edit the same project you use visually. External providers can require an account, network access and payment, and their data handling is separate.

Named client/version/OS setup will accompany qualification. No universal client compatibility or current public installer path is promised here.

Verify the connection with a read.

  1. Open carvercat normally and choose a disposable project.
  2. Use the supported connection setup for the candidate and your named client. If a client is offered under Connect to AI, follow that route; otherwise use its release-specific instructions.
  3. Preserve unrelated client configuration. Reconnect/restart the connector where the setup requires it.
  4. Ask for open comps and current state without editing.
  5. Try one simple object, inspect it, Undo and verify a saved/reopened test copy.

Desktop attachment and headless batch authoring are separate modes. A connector can exist while the visible editor is closed; connection availability is not project readiness.

Use the installed app’s capabilities.

Ask the client to discover supported tools/resources and identify project, comp, object and range through stable identities. Read focused task guidance. Do not hardcode a private schema, assume old command descriptions are current, or select by an unverified position.

A refusal/no-op supplies context. The agent should adjust the plan within supported operations rather than bypass a restriction.

Keep the edit bounded.

  1. Describe the intended change and what must be preserved.
  2. Read state and relevant prerequisites.
  3. Make the change using applicable revision checks.
  4. Inspect structured state and a visible/rendered result.
  5. On conflict, reread and replan.
  6. Stop at the agreed point and describe changes.

Correct values alone do not establish visual quality. Delivery requires its own target/range/method review.

Refresh after manual changes.

Let the person edit the same source. Before the next agent operation, read current state again and preserve those changes. An old transaction or cached wording should not overwrite a newer manual revision.

Recover from disconnects and uncertainty.

If no editor is ready, open it normally and retry a state read. After a timeout, inspect state/history/output before repeating an edit. If sending to Resolve, inspect the actual destination first. A lost response does not prove a failed action.

Local operation is not a blanket privacy guarantee. Share only what the task needs, review your client/provider permissions and redact customer data from reports.