Playwright MCP is Microsoft's official MCP server that gives agents a real browser. Its key idea: the agent sees accessibility-tree snapshots instead of screenshots - structured, deterministic and far more token-efficient than vision-based control. It has become the default way to connect coding agents (Claude Code, Copilot & co.) to a browser, e.g. to verify web UIs they just built.
- Playwright CLI: since mid-2026 Microsoft recommends the CLI variant for coding agents - each action runs as a shell command, artifacts go to disk instead of the model context (~4x fewer tokens). MCP remains the fit for long autonomous browser loops.
- Playwright Test Agents (planner / generator / healer) generate and self-heal Playwright tests - promising, but expect to review everything they produce.
Distinction from Browser Use: Playwright MCP/CLI is a deterministic tool layer that your agent drives - you own the loop and the harness. Browser Use is an opinionated agent library with its own loop and cloud offering, geared towards scraping and extraction.
Team Insights
We use Playwright MCP in coding-agent workflows to let agents verify frontend changes themselves (open page, click through, check the result) - a feedback loop that improves results in the sense of Harness Engineering.
Recommendation
Default choice when an agent needs a browser. Treat visited pages as untrusted input (Prompt Injection Awareness) and prefer the CLI variant for token-sensitive coding-agent setups.