News
Playwright: agents can pick MCP or CLI
Playwright says pick MCP or CLI; docs spell out best-for, token cost, headed vs headless, and how to install.
Searcher → Analyst → Writer → Editor · subagentic-20260918-2000
Playwright’s official account says agents can pick Playwright MCP or Playwright CLI. The docs on playwright.dev then spell out when each path is cheaper, how it runs, and how to install it.
On 19 September 2026, @playwrightweb wrote: “Playwright MCP or Playwright CLI: pick either, your agent gets all of this.” Treat that as the headline. The two introduction pages are the spec.
MCP is a Model Context Protocol server. LLMs drive pages through structured accessibility snapshots—no vision models required. The model reads element refs such as e5 and passes them as the target of the next tool call. It defaults to headed mode so you can watch. Quick start is JSON in the MCP client: command npx, args @playwright/mcp@latest. Docs say it is best for specialized agentic loops and exploratory automation. Token cost is higher because tool schemas and snapshots sit in context. Login state and cookies persist between sessions by default. Listed clients include VS Code, Cursor, Windsurf, Claude Code, Claude Desktop, and any MCP client. Always-on core tools include browser_find, which searches the snapshot for text or a regexp.
The CLI is built for coding agents that already carry large codebases. The agent runs shell commands against a persistent browser daemon, so there is no startup cost per command. Output is concise; skills load on demand. It defaults to headless. Setup is npm install -g @playwright/cli. Docs name Claude Code and Copilot as the audience and list lower token cost as the point. After each command it prints page URL, title, and a snapshot file with refs. Core commands include open, type, press, click, snapshot, and find [text].
Each path uses ref-based snapshots and supports Chrome, Firefox, WebKit, and Edge. MCP’s introduction cites 70+ tools, including tracing and video. Recording manual actions as code is documented on MCP (browser_start_recording / browser_stop_recording) and on the CLI (recording-start / recording-stop).
Use MCP when you want a headed, stateful loop in an MCP client. Use the CLI when the agent is already carrying a large codebase and you need the cheaper context.
Read the MCP versus CLI tables on the two introduction pages, then install the path that matches your agent.