I have gathered comprehensive data. Here is my investigation summary. --- # Playwright MCP: Browser Automation for AI Agents (2026) ## What it is Playwright MCP is Microsoft's Model Context Protocol server that gives AI agents (Claude, Copilot, Cursor, Codex, etc.) direct control of a real browser. Instead of the agent guessing at DOM or screenshots, the server exposes ~40+ tools (the mcp.directory guide counts "over 50") that operate on an **accessibility snapshot** — a distilled, token-efficient representation of the page (~200–400 tokens per snapshot vs. thousands for raw DOM or screenshots). Tools span core automation (navigate, click, fill, drag, upload, evaluate), network (monitor/mock requests), storage state (persistent login), DevTools/CDP, vision (optional), PDF, and testing. ## What changed in 2026 (the big story) **1. The CLI + SKILLS pivot — the most significant shift.** The official GitHub repo and docs now prominently feature a "Playwright MCP vs Playwright CLI" comparison, and the maintainers now **recommend the CLI + SKILLS approach over MCP for coding agents**. Rationale: MCP's tool-call round-trips and snapshot overhead are token-expensive; a CLI (`@playwright/cli`, headless by default) with bundled "skills" (prompt templates) is more token-efficient for agents working in large codebases. MCP remains the recommended path for specialized agentic loops, exploratory automation, and non-coding agents. This is a genuine 2026 repositioning of the project's guidance. **2. Rapid release cadence and feature growth.** The package shipped 10+ versions in the first half of 2026 alone (v0.0.69 Mar 30 → v0.0.78 Jul 9). Notable 2026 additions: - **v0.0.78 (Jul 9, 2026):** new `browser_find` tool (regex/text search of the accessibility snapshot, cheaper than full snapshots), distilled "less verbose" snapshots to cut tokens, new `--mobile` flag and `--device` option for device emulation, and `browser_navigate` now surfaces non-2xx status codes. - **v0.0.77 (Jun 29, 2026):** screenshot `scale` option (CSS vs device pixels), configurable heartbeat timeout, secret redaction in console logs, and `data:` URL payloads excluded from snapshots/network output. - **v0.0.76 (Jun 10, 2026):** `browser_video_show_actions`/`browser_video_hide_actions` (overlay action annotations on recorded video), `--output-max-size` cap, Firefox BiDi channel support. - **v0.0.74 (May 6, 2026):** browser-extension mode now manages multiple tabs. - **v0.0.63 (Feb 3, 2026):** CLI readme updated with skills installation instructions — the CLI+SKILLS story consolidating. **3. A real security incident.** **CVE-2025-9611** (published Jan 7, 2026, CVSS 4.0 = 7.2 HIGH, per VulnCheck CNA): the MCP server before v0.0.40 failed to validate the Origin header, enabling a **DNS-rebinding attack** that could invoke MCP tools from a malicious website. Fixed in v0.0.40. This is the kind of risk that comes with giving agents browser control, and it drove hardening (path-traversal checks on static file serving, regex validation, secret redaction) in later releases. **4. Maturation of the platform.** Node.js requirement moved to 20+; three profile modes (Persistent, Isolated, Browser extension); standalone HTTP server (`npx @playwright/mcp@latest --port 8931`); config-file support; and a broad client ecosystem (VS Code, Cursor, Windsurf, Claude Desktop, Claude Code, Codex, Copilot, Goose, Grok, Junie, Amp, Antigravity, Cline, Factory, Gemini CLI). ## Why it matters - **It made browser automation a first-class agent capability.** Before MCP, agents either scraped text or used brittle screenshot/DOM heuristics. The accessibility-snapshot approach is dramatically cheaper and more reliable, and it's now the de-facto standard for agentic browsing. - **It's a massive adoption signal.** `@playwright/mcp` pulled **~6.16 million downloads in the week of 2026-07-24** (npm API), with 404 published versions and 104 dependents; the GitHub repo has **~35.7k stars / 3k forks**. - **It's actively maintained by Microsoft** (Playwright team), with signed releases, a release-notes workflow, and a dedicated `.claude/skills` workflow in the repo — the team dogfoods its own agent tooling. - **The CLI+SKILLS pivot matters strategically:** it signals that MCP's token overhead is a real cost in production agent loops, and that the ecosystem is converging on hybrid approaches — MCP for rich tool surfaces, CLI+skills for cheap, high-volume coding-agent work. ## Strongest sources (verified live) 1. **Official GitHub repo — microsoft/playwright-mcp** (35.7k stars; the MCP-vs-CLI recommendation; client list; Apache-2.0). *Primary source.* 2. **Official docs — playwright.dev/docs/getting-started-mcp** (40+ tools, token counts, three profile modes, standalone HTTP server, config file, Node 20+). *Primary source.* 3. **Official docs — playwright.dev/mcp/introduction** (MCP vs CLI comparison table; new tools like `browser_generate_locator`, `browser_verify_element_visible`, tracing/video/PDF tools). *Primary source.* 4. **Official docs — playwright.dev/docs/getting-started-cli** (the 2026 CLI+SKILLS path for coding agents). *Primary source.* 5. **GitHub Releases/Tags** (v0.0.78 Jul 9, 2026; v0.0.77; v0.0.76; v0.0.74; v0.0.63 — concrete feature/fix history). *Primary source.* 6. **npm API** (`api.npmjs.org/downloads/point/last-week/@playwright/mcp`): 6,164,276 downloads, 2026-07-24→30. *Primary metric.* 7. **NVD — CVE-2025-9611** (CSRF/DNS-rebinding flaw, CVSS 7.2 HIGH, fixed in 0.0.40). *Primary security source.* 8. **mcp.directory guide** (the starting point; dated 2026-05-11; good practitioner overview of installs, recipes, and troubleshooting — though its "50+ tools" count is slightly higher than the official "40+"). ## Gaps / notes - I could not pin down the **exact original launch date** (the earliest tags I reached were v0.0.59–v0.0.68, Jan–Feb 2026; the very first release and the original 2025 announcement blog post weren't captured — one secondary source dates the initial release to March 2025, another to mid-2025; I'd verify against the earliest GitHub tag before citing a specific launch date). - The mcp.directory guide's tool count ("over 50") differs slightly from the official "40+" — the official number is the safer citation. - I did not capture the exact npm **weekly-downloads trend** over time (only the single latest week), so the growth trajectory is inferred from release cadence and the 6.16M weekly figure.