136 / 139 · 13 Browser Automation with Playwright · AI-Assisted Testing← prev ⊞ all next →☰ Read as one page

16.5How Coding Agents Drive the Browser: CLI and MCP

Test generation is only useful if the model can see the page. Two official integrations give AI agents live browser access — and as of July 2026 they are not equal.

Playwright CLI: The Recommended Path for Coding Agents

Microsoft now recommends the Playwright CLI (@playwright/cli on npm) over the MCP server for coding agents — it uses roughly 4x fewer tokens (~27k vs ~114k for a typical task). Instead of holding a long-lived tool session inside the model's context, the agent shells out to short commands:

npx playwright-cli open https://example.com
npx playwright-cli snapshot                    # page snapshot with element refs
npx playwright-cli fill e8 "user@example.com"  # act on elements by ref
npx playwright-cli click e12

Two design choices keep the context small:

Each snapshot assigns short element refs (e8, e12) that the agent uses in follow-up commands — no verbose selectors round-tripping through the model
Snapshots are saved to disk as YAML under .playwright-cli/ rather than dumped into the conversation; the agent reads back only what it needs

Running playwright-cli install --skills generates a SKILL.md file that teaches skill-aware agents (Claude Code, Cursor, Codex CLI, and others) how to use the CLI without you writing instructions by hand.

Playwright MCP

MCP (Model Context Protocol) is an open protocol that lets AI agents interact with tools — including Playwright — through a standardized interface. With Playwright's MCP server (published to the official MCP Registry each release), an AI agent can browse the web, interact with pages, and extract information programmatically.

AI Agent  --(MCP Protocol)-->  Playwright MCP Server  --(Playwright API)-->  Browser

The AI agent sends high-level commands ("navigate to the login page", "fill in the email field"), and the MCP server translates them into Playwright actions.

Rule of thumb: for coding agents (Claude Code, Codex CLI, and similar), prefer the CLI — same control, a fraction of the tokens. MCP remains the right fit for MCP-native clients and chat-style assistants that cannot shell out to commands.

Practical Applications

AI-powered exploratory testing: An agent navigates the app, tries different paths, and reports anomalies
Test data setup: AI agent creates test scenarios through the UI when API is not available
Accessibility auditing: Agent crawls pages and reports accessibility issues
Visual regression triage: AI reviews screenshot diffs and classifies them as intentional or buggy

Current Limitations

Agentic browser control has matured quickly, but as of July 2026:

Agents are still slow compared to scripted tests
They make mistakes (wrong locators, incorrect assertions)
Cost per test run is higher than traditional automation
Best suited for exploration, authoring, and healing — not for executing continuous regression suites, where deterministic scripted tests remain the right tool