Docs

Automate Stunt Double from Claude

Once the MCP server is connected, you can run most of your workspace from a conversation instead of clicking through the dashboard. This tutorial uses Claude, but every MCP client works the same way.

Before you start

Connect your client and sign in: pick it from the client pages (Claude, Claude Code, ChatGPT, Cursor and more). Then ask "List my Stunt Double workspaces." If your workspaces come back, you're ready.

What you can ask for

Ask in plain language. Claude turns each request into tool calls on actors, checklists, workflows, interviews and feedback, and reads back the result:

  • "Run the checklist for our signup flow and tell me if anything failed."
  • "Create an actor for a first-time free-trial user and give it our brand guidelines to check against."
  • "Launch a new round of our onboarding interview and summarize the top themes."
  • "List the workflows that ran in the last day and flag any failures."
  • "Summarize the open feedback on the pricing page and dismiss the duplicates."
  • "Run the checkout checklist against this preview and comment the result on the pull request."
  • "How well can AI agents use example.com?" For a site the Index does not track yet, Claude can request a score.

Name the workspace or project when you have more than one, and Claude searches before it creates, so it reuses the actor or checklist that already covers a flow.

Start from a recipe

For the common jobs you don't have to write the request yourself:

  • Prompts. The server ships ready-made prompts such as verify_change, run_user_research, triage_feedback and setup_guardrails. Each client lists them in its own way; the client pages show where.
  • Skills. The skills go further: each one teaches the agent a full job, such as verifying a pull request preview or standing up guardrails. The Claude Code and Cursor plugins install them for you.

Let runs finish

Checklist runs, workflow runs and interviews take a few minutes. Claude starts the run, then checks back every 30 to 60 seconds until it finishes, so you can ask for something else in the meantime. In Claude and ChatGPT, the run comes back as a card with the verdict, the score and each check's evidence, and the card keeps itself up to date while the run is going.

Give each client only what it needs

A connection that only reports on results needs mcp:read, nothing more. Approve mcp:write for a client that should build checklists and actors, and mcp:run only where it should start runs, since runs use your workspace's run allowance. Connect explains how scopes hide tools and how a client without mcp:run handles automations.

Next steps

On this page