MCP (Model Context Protocol) is an open standard for giving an AI agent tools: each MCP server exposes actions such as "open this page", "read this file" or "query this CI run", and the agent decides when to call them. For test maintenance, the useful servers are a browser (Microsoft's Playwright MCP server lets an agent drive a real browser through the accessibility tree), your repository, and your CI or issue tracker.

A realistic workflow today looks like this:

1. A test fails in CI. A developer or a scheduled job hands the agent the failing test, its error and trace, and the recent diff.
2. The agent reproduces the failure in a real browser through the Playwright MCP server, inspects the page, and proposes a fix to the test or the page object. Playwright's own test agents (planner, generator and healer) package this loop.
3. The agent re-runs the test and opens a pull request. A person reviews it like any other change.

Keep the controls tight. The agent should never merge on its own or edit assertions to make a test pass, because a "healed" test can hide a real regression. Give each MCP server the narrowest permissions it needs, run it against test environments only, and keep secrets out of prompts and tool output. Track how often each test needs repair: a test that keeps drifting points at an unstable locator or a missing test id.

Sources: Model Context Protocol, Playwright MCP, Playwright test agents