Tool ยท in the coding agent
Browser access for the agent
Gives the agent itself a browser to act in, through the page's accessibility tree, with the console and network in view.
No illustration yet
catalog/tools/agent-browser-access.svg This tool gives the coding agent a browser of its own to act in while it works. It sees the page through the accessibility tree, with the console and the network in view.
Finding it
It is added to the coding agent as an MCP server, in the agent's MCP configuration: a project file shared with the
team (.mcp.json in Claude Code) or the user's own settings. The agent's list of connected servers (/mcp in Claude
Code) shows whether it is there and running.
Using it
- Explore before writing a test: open the app, take a snapshot of the accessibility tree, act on elements by role and name, and note what a test should check.
- Read the console and the network directly: script errors and failed requests come from the browser, not from a guess about the page.
- Take a screenshot when the result has to be seen, not only read.
- Start from a known state: an isolated browser profile for each session, so logins and caches left by earlier runs do not change what it sees.
Examples: Playwright MCP, Chrome DevTools MCP.
Built from catalog/tools/agent-browser-access.md.