AI Agent tool¶
The module registers one AI Agent tool (an ai / ai_agents function call).
It gives a Drupal AI Agent a look at a live rendered page — much as the
Playwright MCP does for an editor — but from inside Drupal, as an agent tool the
Drupal Canvas AI assistant can call.
Browser preview (Playwright)¶
| Plugin id | ai_playwright:browser_preview |
| Function name | ai_playwright_browser_preview |
| Group | information_tools |
| Permission | Use the AI Playwright browser tool (use ai playwright), or administer ai agents / use Drupal Canvas AI |
Inputs¶
Both are optional:
| Input | Description |
|---|---|
url |
A path on this site (e.g. / or /about). Empty means the front page. |
task |
A note of what to check on the page, echoed back for context. |
Output¶
One text block with:
- the URL that was opened,
- the page title,
- a screenshot saved as a managed file — its id can be passed to
describe_imageso the agent actually sees the page, - any console errors, and
- the page's visible text (trimmed).
This is the same signal the Playwright MCP surfaces from a snapshot — a screenshot, the page text, and console errors — delivered to the agent in a single call so it can verify what it built.
Use it with the Canvas AI assistant¶
Add Browser preview (Playwright) to the Canvas AI assistant's tool list
(Configuration → AI → Agents → your orchestrator → Select tools). Then ask the
assistant to open, check, or verify a page. It calls the tool, gets the
screenshot file id, calls describe_image to see the render, and answers or
refines the build.
Scope
This first release captures a page (navigate → screenshot → text → console errors), which is what an agent needs to look at a result. Interactive steps (click, type, multi-step flows) are out of scope for now.