MCP test automation explained
In 2026 there are roughly three ways to put AI in the loop on E2E testing. Two of them have a quiet cost problem. This is a short, opinionated explainer of where MCP fits.
What MCP actually is
MCP — the Model Context Protocol — is the open standard Anthropic shipped for connecting LLMs to external tools. It lets Claude (and Cursor, Continue.dev, Cline, and any other compatible client) call native verbs on a service through a single transport.
Before MCP, plugging Claude into a testing platform meant either a custom Anthropic API integration or wiring up Computer Use. After MCP, it's a URL plus OAuth.
The three flavours of "AI testing" in 2026
1. Agent-driven (Claude Computer Use, raw playwright-mcp)
The LLM drives a browser visually on every run. Brilliant at exploration. The catch: you pay LLM tokens per step, on every replay. For a 100-step test run 100× a year, that's ~$320 in tokens alone — before CI compute. Adoption is real for one-off automation; it doesn't scale to a regression suite.
2. Hand-written code (Cypress, Playwright)
Engineers write the test code. Fast and deterministic at runtime. Zero LLM tokens. The catch is on the human side: hours of code per flow, manual fixes on UI churn, abandoned suites when the original author leaves.
3. MCP test automation (Claude + a platform)
The split: Claude designs the test once via MCP, the platform stores it as a Playwright-grade recording, replay is deterministic. LLM is invoked only when self-healing actually needs to run. You get the authoring speed of (1) and the runtime cost of (2).
Why this matters now
Two reasons it's a 2026 story specifically:
- MCP went mainstream — Anthropic opened it, every major AI client adopted it, the SDK crossed 90M+ monthly downloads. The standard exists.
- The cost gap widened — Computer Use is more capable than ever, but its per-step token cost moved the conversation from "is this nice?" to "is this billable?". Teams running E2E nightly + per-PR can't afford agent-driven replay.
Where MCP test automation does NOT fit
Honest scoping:
- Pure exploratory — if your goal is "let the AI poke around and report what broke", Computer Use is still the right tool. MCP shines when you want repeatable assertions.
- Non-browser flows — desktop apps, native mobile. (Maestro MCP covers mobile; we focus on web.)
- Proprietary visual diff — for pixel-perfect visual regression, a dedicated tool like Applitools is still better.
If you're evaluating
Three questions worth asking any "AI testing" vendor:
- What's the LLM cost per replay? (Should be near zero for 90%+ of runs.)
- Where do tests live, and do they run without the AI?
- What does self-heal actually patch — the failing step only, or the entire flow?
Read the cost math on the main page for the numbers, or try the ROI calculator with your team's actual numbers.