Playwright MCP for agent browser tools

How it works

The server gives an MCP client browser actions and can expose structured accessibility snapshots. A client can choose an action based on the page structure and inspect the updated state.

What you need

Use a supported MCP client and set an explicit browser profile policy. The server exposes actions to that client, which must decide what the agent may call.

What to try first

Connect it to an MCP client and interact with a local fixture page. Test an element with an accessible label and another with ambiguous structure to see what the model receives.

Change an accessible name while leaving screen layout untouched. The resulting snapshot helps show how much an automation depends on semantic labels.

Where to look for failures

Browser tools can operate with active credentials. Use isolated profiles for evaluation and enforce action permissions at the client and browser boundaries.

Related Franklin pages

Hold the browser task constant and compare the observation and action layers. An explicit selector, a structured browser snapshot and a model-chosen action fail in different ways.

• Let a model navigate a browser and act on page elements. Browser Use for model-driven web tasks
• Combine explicit browser steps with natural-language actions, extraction and observation. Stagehand for code and model-assisted browser automation

For an agent interface that reaches mobile apps, see Franklin’s guide to Phone-Use. Phone-Use

Project documentation

• Current project README
• README at the reviewed revision