Playwright CLI vs MCP debate heats up

Microsoft’s own Playwright docs now draw a sharper line: use Playwright CLI for coding agents, and use Playwright MCP when an agent needs a persistent, stateful browser session. (playwright.dev) (github.com) The distinction starts with how each tool talks to a page. Playwright MCP exposes browser actions through a Model Context Protocol server and returns structured accessibility snapshots, while Playwright CLI runs discrete commands and keeps responses shorter. (playwright.dev 1) (playwright.dev 2) Microsoft’s GitHub readmes now say that plainly. The Playwright MCP repo says coding-agent users “might benefit from” CLI plus skills, and the Playwright CLI repo says coding agents are “the best fit” for the CLI. (github.com 1) (github.com 2) A browser accessibility tree is the machine-readable outline of a page — buttons, headings, text boxes, and labels, not pixels. Playwright MCP uses that tree so an agent can reason over page structure without relying on screenshots or a vision model. (playwright.dev) The CLI takes a different route. Playwright describes it as a command-line interface for coding agents with “token-efficient” commands and installable skills that teach agents which commands and workflows to use. (playwright.dev 1) (playwright.dev 2) That matters because coding agents are already carrying prompts, code, test output, and repository context in the same window. Microsoft’s guidance says CLI workflows avoid loading large tool schemas and verbose accessibility trees into that context. (github.com) (github.com) Playwright is also building more agent behavior into the test stack itself. Its test-agents docs now ship three built-in agents — planner, generator, and healer — with the healer described as the component that executes a test suite and automatically repairs failing tests. (playwright.dev) That is where the current argument in quality-assurance circles is landing in practice. Teams are separating two jobs that used to blur together: deterministic end-to-end checks that must stay stable in continuous integration, and agent-driven exploration or repair work that can tolerate more guesswork. (playwright.dev) (playwright.dev) Playwright’s own best-practices guide still pushes tests toward user-visible behavior and away from implementation details. In plain terms, that means resilient selectors and explicit assertions remain the baseline even as agent tools generate code, inspect pages, or patch broken flows. (playwright.dev) The result is less a winner-take-all fight than a workflow split. Playwright’s latest docs now position CLI as the default control surface for coding agents, and MCP as the heavier tool for deep page reasoning, long sessions, and self-healing loops. (playwright.dev) (playwright.dev)