MCP Playwright problems and fixes

Decide whether connecting, starting a browser, navigating, acting on the page, reading a resource or stopping failed. Review console logs, network details, screenshots and saved state before sharing them; they can contain private information.

The service keeps its own tools visible when browser integration discovery fails, so page automation is not ready yet. Confirm the browser integration is installed in the service's launch environment, then restart the service; discovery runs at startup.

A missing executable points at the integration path; not supported means the engine is not enabled; an invalid type means the name is not recognized. Startup allows 30 seconds (MCP_BROWSER_START_TIMEOUT) within the 120-second tool limit (MCP_TOOL_TIMEOUT). Stop a failed entry from browser_list and start a new browser.

Ready does not prove the initial URL loaded: call browser_navigate and take a snapshot. Without a per-browser HEADLESS setting, the service picks headless or headed from its own display. Grant only the permissions the task needs in the start configuration.

Use the same MCP session: separate cli invocations do not share browsers, and a stop or expired session invalidates an ID. List browsers through the current connection, or start a new one and sign in again.

Read summary.errors and every results entry. A member can fail after others finished, so inspect it directly and avoid repeating completed side effects. Snapshot references are per page and per engine.

Wait for the retry delay in a rate-limit error. A timeout does not cancel a browser action, so inspect before repeating a submission. Keep actions on one browser sequential; use groups for concurrency across browsers.

Screenshots are published as resources: read the returned URI through the same session, before it ends. Group screenshots are not published per member; take direct screenshots instead. Stop each browser before ending a session so its video is finalized. Action recording publishes a codegen resource only after at least one recorded action.

GPU selection follows the supplied rendering environment. For a software comparison, start a separate browser with PLAYWRIGHT_MCP_GPU=off in its env. A reported renderer name is not proof of acceleration; check the page's actual outcome.

The service expected hardware rendering, but the browser fell back to CPU rendering such as SwiftShader or llvmpipe. Remove a forced hardware gpu value or report the environment's graphics access to its maintainer; to continue on CPU, start with PLAYWRIGHT_MCP_GPU=off.

Check the service log entry naming the Apple Chromium backend. Export any evidence you need, set PLAYWRIGHT_MCP_APPLE_ANGLE=gl-egl in the env map of browser_start for that browser (or in the service's launch environment for every browser), start a new Chromium browser and check the page visually: this backend is not yet confirmed on such hosts. If Chromium stays blank, use Firefox for the visual check and report the backend and symptom to the environment maintainer.

Input paths are on the service host. Saved browser state is sensitive and can expire, so confirm the signed-in account before protected actions. Use the environment's trusted certificates rather than bypassing certificate errors.