Tools Reference

A comprehensive guide to the MCP tools exposed by the Grip Daemon, optimized for AI agent workflows.

Overview

The Grip Daemon provides a standardized suite of MCP tools designed to let AI agents interact with web pages reliably. These tools translate high-level requests into low-level Chrome DevTools Protocol commands.

Important: Element references generated by the snapshot tool are ephemeral. They expire upon navigation or significant DOM updates. Always request a fresh snapshot before interacting with a changed page.

If your AI tool is not yet connected to the daemon, refer to the Configuration Guide.

Tools Reference

snapshot()

Read the page accessibility tree and get stable element references.

When to use: After every page load or major DOM mutation.

highlight()

Draw a colored outline around a specific element reference.

When to use: To visually confirm the correct target before clicking or typing.

click()

Click an interactive element (button, link, toggle).

When to use: Submitting forms or navigating through UI flows.

fill()

Input text into a field or textarea.

When to use: Logging in, searching, or completing forms.

read_logs()

Extract the browser's console output (errors, warnings, logs).

When to use: Verifying if a recent action caused a client-side exception.

read_network()

Retrieve recent network requests (XHR/Fetch) in HAR format.

When to use: Debugging failing API calls or monitoring payload data.

screenshot()

Capture the full viewport or a specific element as a base64 image.

When to use: Visual verification or sharing exact state with the user.

pick_element()

Activate the visual picker in the browser.

When to use: Requesting human-in-the-loop intervention for complex element selection.

navigate()

Transition the browser to a new URL.

When to use: Starting a new task or flow.

eval()

Execute raw JavaScript within the active page context.

When to use: For advanced edge cases not covered by standard tools.

Optimal Workflow

For the highest reliability, AI agents should adhere to this interaction loop:

  1. Inspect: Call snapshot() to retrieve the current layout and available interactive elements.
  2. Verify: (Optional) Call highlight(ref) to ensure the intended target is correct.
  3. Act: Call click(ref) or fill(ref, value) to manipulate the page.
  4. Validate: Call read_logs() to verify no unexpected errors were triggered by the action.