Feature: Browser automation via MCP (Playwright)

[Rustam-Z]

Goal

Enable the bot to drive a real browser — fill forms, click buttons, extract data from JS-heavy pages, take screenshots, automate web flows.

Why

Currently the bot can only do read-only HTTP fetches via WebFetch / WebSearch. That misses:

• JS-rendered content (SPAs, dashboards, infinite scroll)
• Authenticated flows (logging in, navigating gated content)
• Form submissions (filling and submitting on user's behalf)
• Multi-step UI flows (booking, checkout, account management)
• Visual extraction (screenshots, layout inspection)

OpenClaw has this natively. Adding it to pyclaudir closes a major capability gap for business automation use cases.

Approach

Use a Playwright-based MCP server. Two options:

A) Existing MCP — there are community Playwright MCP servers (e.g. @modelcontextprotocol/server-playwright). Pull in, configure in plugins.json, exposed as tools. Lowest effort, fastest path.

B) Custom MCP wrapper — write a thin pyclaudir-specific MCP server that wraps Playwright. More control over what gets exposed, easier to lock down dangerous actions.

Recommend A for v1, B if security or scope becomes a concern.

What the bot would gain

New tools:

• browser_navigate(url)
• browser_click(selector)
• browser_fill(selector, value)
• browser_screenshot() — returns image
• browser_extract(selector) — returns text
• browser_wait_for(selector, timeout)

Risks / safeguards

• Token cost: browser automation runs longer than HTTP — each session can burn many tool calls. Keep in plugins.json opt-in like the bash tool group.
• Privacy: sites should not run with the operator's logged-in browser profile by default — use isolated browser contexts.
• Abuse: non-allowlisted users could try to weaponize for scraping; access control already handles this.

Acceptance

• Browser MCP enabled in plugins.json
• Bot can navigate, click, fill, extract, screenshot
• Headless mode by default
• Off by default (in line with bash/code tool groups)