Kiwi

Kiwi for Playwright

Have the agent on a Kiwi step drive a browser through Playwright MCP to check a change in the running app and capture what it saw.

  • Your agent works in Playwright through its own connection. Kiwi does not call Playwright and never holds its credentials.
  • A workflow skill can list Playwright among the tools it requires. When you supply your agent's available tools to a skill check and Playwright is not among them, the check reports SK15-UNSUPPORTED-TOOL.
  • A project Admin can add links to Playwright, with notes, under Connections and bind them to the steps that use them. The agent reads them with list_connections.
  • The agent can report the tools and plugins it ran with, including its Playwright connection, with report_configuration.
  • The agent submits links to its work in Playwright as the step's outputs. People review them, and sign-off stays with the people your workflow names.

What your agent can do in Playwright

  • Navigate pages and act on elements from accessibility snapshots rather than screenshots.
  • Click, type and fill forms, then verify that the expected elements and text are visible.
  • Read console messages and network requests to explain a failure.
  • Take screenshots, save a page as PDF, and record traces or video when they are enabled.
  • Generate locators for elements so the agent can write a Playwright test.

Set up

  1. Connect your agent to Kiwi

    Add Kiwi's hosted MCP server, https://kiwiflow.io/mcp, to your coding agent.

  2. Connect your agent to Playwright

    Run the Playwright MCP server where your agent runs. The credentials stay with your agent.

    npx @playwright/mcp@latest
  3. Example skill

    Publish instructions like these as a workflow skill that lists Playwright among its required tools.

    Start from the preview URL and the user flow named in the assignment.
    Use Playwright MCP to walk the flow and verify each expected element.
    Save a screenshot of each checked state and note any console errors.
    Submit the screenshots and a pass or fail note per state as outputs.
    If the preview does not load, stop and report a blocker to the step owner;
    the assigned reviewer decides whether the flow is ready.