What it is
Stagehand is an open-source SDK, maintained by the cloud-browser company Browserbase, for scripts and agents that drive a real Chrome browser. Instead of handing a model the whole task, it lets developers write ordinary steps (goto, click, locator) and call a model only where a page is unpredictable: act performs an instruction such as 'click the sign in button', observe returns the selectors it would use, and extract returns data validated against a schema.
Version 4, released on 10 August 2026, rebuilt it as one driver across TypeScript, Python and Go that talks to Chrome over the DevTools Protocol with no Playwright dependency, and runs part of its logic as an extension inside the browser. 4.3.0 shipped on 20 October 2026; the 3.x line still receives patches. The code is MIT-licensed.
It runs against a local Chrome or a Browserbase cloud browser. Two features, a model gateway that picks the model per step and a server-side cache of repeated actions, work only on Browserbase.
Our review
Stagehand's case rests on a design choice: the model is a tool inside a script, not the pilot. A login flow can find the email field with observe, fill it from an environment variable in plain code, and use act only for the button whose markup keeps changing. Because observe returns real selectors, passwords never pass through the model, which is a better security default than handing an agent the whole form. Browser Use, the closest open alternative, starts from the opposite end, with a model choosing every step.
Version 4 is close to a new product. It dropped Playwright, ships equal SDKs for TypeScript (Node 22.18 or later), Python 3.11 and Go 1.26, and moved part of the runtime into a browser extension to cut round trips. The project has the repository's activity to match: 25,791 stars, about 104 open issues, and four minor releases between 10 August and 20 October 2026. Teams upgrading from 3.x should treat it as a migration rather than a version bump.
The open-source story has a commercial edge. On a local browser every act, observe and extract call goes to the model you configure, from five built-in providers (OpenAI, Anthropic, Google, Groq, Cerebras) or your own callback; there is no base-URL setting, so a self-hosted model needs that callback. On Browserbase, caching is on by default and repeated calls return with no tokens spent, and the Model Gateway bills at provider prices with no markup, according to the docs. Neither works locally, so the cheapest way to run Stagehand at volume is on its maintainer's cloud.
On privacy, the SDK is cleaner than its main rival. In our reading of the repository, the bundled CLI sends PostHog usage events (command names, versions, durations and exit codes, never raw arguments) unless DO_NOT_TRACK or BROWSERBASE_TELEMETRY_DISABLED is set, while the SDK packages carry OpenTelemetry tracing hooks and no PostHog calls. Browser Use, by contrast, ships with telemetry on that includes task text and URLs.
Where it earns its place
A finance-operations engineer at a 50-person SaaS company
Pull invoices every night from supplier portals that have no API
Log in with plain locator code and stored credentials, use act for navigation that breaks on redesigns, and extract each invoice table into a Zod or Pydantic schema. Locally the cost is model tokens per AI step; on Browserbase repeated steps are served from the cache.
A QA lead with a large Playwright suite that breaks on every redesign
Make the flakiest end-to-end tests survive UI changes
Port the suite with the Playwright migration guide and replace only the fragile selectors with act or observe calls. The rest stays deterministic, so model spend is limited to the steps that used to fail.
A product manager budgeting an agent that books appointments for customers
Estimate the monthly cost of 1,000 browser runs a day
At three minutes a run that is about 1,500 browser hours a month: Browserbase Startup at $99 plus 1,000 extra hours at $0.10, about $199, before model tokens and proxy traffic at $10 per GB beyond 5 GB.
Before you commit
- Setup
- pnpm add @browserbasehq/stagehand with Zod 4, pip install stagehand, or go get the v4 Go module; local runs need Chrome installed. The quickstart goes from an empty directory to a working act, observe and extract script in three steps.
- Integrations
- Built-in OpenAI, Anthropic, Google, Groq and Cerebras providers, or a callback for anything else; documented guides for CrewAI, Mastra, LangChain Deep Agents, Vercel AI SDK, Claude Code, Codex and Cursor; Claude and OpenAI computer-use toolsets; WebMCP; a hosted Browserbase MCP server.
- Data and privacy
- Locally, page content goes only to the model provider you configure. On Browserbase the cloud sees every page and form value; session data is kept 7 days on Free and 30 days on paid plans, and HIPAA BAA and DPA are listed for Scale only. The CLI's PostHog telemetry is off with DO_NOT_TRACK=1. No published security advisories.
- Cost at scale
- Free locally apart from model tokens. On Browserbase, Developer is $20 for 100 hours then $0.12 an hour, Startup $99 for 500 hours then $0.10, with proxies at $12 or $10 per GB over the allowance; cached actions cost no tokens.
- Limits
- Concurrency on Browserbase is 3, 25, 100 or 250+ browsers by plan. The gateway rejects stopSequences and has no local fallback. The most-requested open issue is support for the latest Vercel AI SDK model interface.
What's good
- Mix deterministic code and AI steps in one script
- observe returns selectors, so secrets stay out of prompts
- Same API in TypeScript, Python and Go
- No Playwright dependency since v4
What's not
- Caching and the model gateway need Browserbase
- No base-URL option for custom or local model endpoints
- v4 is a breaking rewrite of the 3.x API
- CLI telemetry on by default
Best for
Engineering teams that want AI help on the brittle parts of browser automation while keeping the rest as ordinary, testable code.
Alternatives
Stagehand: common questions
- Is Stagehand worth it?
- Stagehand suits engineers who want to decide, line by line, how much of a browser script a model controls, and it quietly rewards running that script on Browserbase. We rate it 4.3 out of 5.
- How much does Stagehand cost?
- SDK free under MIT. Browserbase browsers: Free 1 hour; Developer $20/month (100 hours, then $0.12/hour); Startup $99/month (500 hours, then $0.10/hour); Scale custom. Model Gateway billed at provider prices.
- What are the downsides of Stagehand?
- Caching and the model gateway need Browserbase; no base-URL option for custom or local model endpoints; v4 is a breaking rewrite of the 3.x API; cLI telemetry on by default.
- Who is Stagehand best for?
- Engineering teams that want AI help on the brittle parts of browser automation while keeping the rest as ordinary, testable code.
- How much does Stagehand cost for a team or in production?
- Free locally apart from model tokens. On Browserbase, Developer is $20 for 100 hours then $0.12 an hour, Startup $99 for 500 hours then $0.10, with proxies at $12 or $10 per GB over the allowance; cached actions cost no tokens.
- Is Stagehand safe for company data?
- Locally, page content goes only to the model provider you configure. On Browserbase the cloud sees every page and form value; session data is kept 7 days on Free and 30 days on paid plans, and HIPAA BAA and DPA are listed for Scale only. The CLI's PostHog telemetry is off with DO_NOT_TRACK=1. No published security advisories.
- What are the best Stagehand alternatives?
- The closest alternatives we have reviewed are Browser Use, Crawl4AI and Firecrawl.
Answers checked Oct 23, 2026.
Make Stagehand? Show this rating on your site
