By name: browser-tooling description: Pick the right browser/UI tool for the job — routes between Playwright tests (deterministic), the agent-browser CLI (default for agent-driven browsing), and pinchtab (human-in-the-loop authed sessions). Use when user reports a UI bug, asks to verify a deployed page, wants E2E coverage, says "look at this page", "did the deploy work", "check the UI", "automate the browser", or "drive Chrome".
Browser Tooling Router
These tools sit at different layers of the stack, not in competition:
AI Agent → control layer (agent-browser / pinchtab) → CDP → Chrome
automation framework (Playwright) ──────────┘
- Playwright — automation framework. You (or a test) write code that says exactly what to do. Most reliable; the testing/scraping/production foundation. Not agent-native (the LLM must emit Playwright code).
- agent-browser (Vercel Labs) — agent-native browser CLI. Compact text/a11y representations,
agent-browser open … / click @e2 / fill @e3 "…". Token-efficient, fast, shells out — no MCP context tax. - pinchtab — agent-oriented control server: a persistent, logged-in, watchable browser over a lightweight HTTP API + accessibility snapshots. Best when a human watches/intervenes on an authed site.
The stack we run (your endorsement): Playwright + agent-browser for ~90% of work; pinchtab on standby for human-in-the-loop authed sessions. Chrome DevTools is not an explicitly-supported standing tool, and there are no browser MCP servers loaded — the CLI layer gives the same agent-native control (a11y snapshots, persistent session) without paying a per-session context tax.
Decision tree
"Look at this page" / "Did the deploy work?" / smoke check / drive a flow → agent-browser CLI (default)
agent-browser open <url>, thenclick @e2/fill @e3 "…"/ snapshot. ~200–400 tokens/page.- First choice for almost everything an agent does in a browser.
Human-in-the-loop on a logged-in site (you watch, agent drives) → pinchtab (standby)
pinchtab serve --port 9867thencurl localhost:9867/snapshot. ~800 tokens/page.- Reach for it only when agent-browser's headless model isn't enough — i.e. you need a persistent, visible, authenticated browser.
"Catch this regression forever" / deterministic scrape / production workflow → Playwright tests
- Free per run. Write tests in the project's
web/tests/e2e/. Use Daily fake-media flags for WebRTC. - After root-causing a bug with the CLI, write a Playwright test so it can never silently regress.
"Long agentic flow across 20 screens where selectors rot" → Stagehand (per-project)
@browserbasehq/stagehand— natural-languageact/extract/observe. Costs LLM tokens per run. Install per-project, not global.
"Why is this slow / what error fired / network looks wrong" → no standing tool
- We dropped the Chrome DevTools MCP (overlap + context tax). If you genuinely need perf/network/console traces, launch
npx chrome-devtools-mcp@latestad-hoc for that one session, then drop it — don't make it standing.
Token budget reference
| Tool | Layer | Cost shape |
|---|---|---|
| Playwright tests in CI | framework | 0 / run (one-time write cost) |
| agent-browser CLI | agent control | ~200–400 tokens / page · no MCP tax |
| pinchtab CLI/server | agent control | ~800 tokens / page · no MCP tax |
| Stagehand | agent framework | LLM tokens / test run |
| dropped | was ~13.7k always-on context | |
| dropped | was ~18k always-on context · launch ad-hoc if ever needed |
Why CLI, not MCP, for browser control
An MCP server taxes every session's context with its tool schemas whether or not you browse. The agent-native value of Playwright-MCP / pinchtab-MCP — structured a11y snapshots + a persistent browser session — is exactly what agent-browser (CLI) and pinchtab (server) already give you on demand, at zero standing cost. Note Playwright-the-framework (tests) is a different layer from Playwright-MCP (agent control); we keep the former, drop the latter. So nothing is lost by removing the browser MCPs.
Workflow patterns
Bug report → permanent test:
- Reproduce/poke with
agent-browser(or pinchtab if you need to watch a logged-in session). - Once root-caused, write a Playwright test so it's protected forever.
WebRTC / Daily.co: Playwright tests — --use-fake-device-for-media-stream, --use-fake-ui-for-media-stream, --use-file-for-fake-audio-capture=path.wav, --use-file-for-fake-video-capture=path.y4m.
See also
- Full guide with examples:
prompts/guides/browser-tooling.md playwright-e2e-testingskill — writing the actual deterministic tests