Feature : Expose a console panel in Playwright MCP

[srijan2607]

Summary

The MCP server only returns console output as plain strings (via browser_console_messages and the console updates in snapshots). When pages log objects or emit multiple messages, MCP clients cannot see structured arguments, timestamps, or clear/limit the history. This makes debugging from agents harder than using the DevTools console.

Steps to reproduce

1. Start the MCP server (e.g., npx @playwright/mcp --headless).
2. Open a page that logs objects and errors:

   <script>
     console.log('hi', { foo: 1 });
     console.error('oops');
   </script>

3. Call browser_console_messages.
4. Observe that the output is flattened strings (object values lost) and there is no way to clear or paginate the console log.

Expected behaviour

• Ability to inspect console entries similar to the DevTools console: level, timestamp, source location, and serialized arguments.
• Option to limit how many entries are returned and clear the stored history after reading so agents can poll cleanly.

Proposal

• Add a new MCP tool browser_console_panel (core capability) that:
  • accepts level (error|warning|info|debug), limit (N most recent), includeArgs, includeTimestamp, and clear flags;
  • returns console entries formatted with level, ISO timestamp (optional), source URL/line, and serialized arguments/stack traces;
  • optionally clears stored console messages when clear is true.
• Keep browser_console_messages unchanged for backward compatibility.
• Introduce a gated DevTools data capability (e.g., --caps=devtools) that exposes CDP-backed tools without driving the DevTools UI:
  • browser_network_entries (read-only): recent requests/responses with URL, method, status, timing, headers, and size; bounded by limit and optional URL filter.
  • browser_trace_capture (read-only, opt-in): start/stop a performance trace and return/save the trace JSON; optional categories and duration.
  • browser_debugger_info (read-only): list loaded scripts and recent error stack traces; optionally capture a snapshot of active breakpoints later.
  • browser_coverage_snapshot (read-only): return JS/CSS coverage percentages and per-script counts (if feasible via CDP).

[srijan2607]

@pavelfeldman what do you think ??

[johnonline35]

I love this idea and added this feature request to verdex issues fyi: verdexhq/verdex-mcp#7