Stagehand
Also known as: Stagehand SDK, stagehand.dev, Browserbase Stagehand
MIT licensed open source browser automation framework giving act, extract, observe and agent primitives, letting developers mix natural language and deterministic code, with action preview, caching and MCP access from any model.
Stagehand is an MIT licensed open source browser automation framework built by Browserbase, and it occupies a deliberate middle position between two failure modes. Traditional frameworks like Playwright and Selenium execute explicit commands but break whenever a site changes a selector. Full agent products take a prompt and act autonomously, at the cost of predictability. Stagehand lets a developer choose which behaviour applies where.
It exposes four primitives: act, extract, observe and agent. The first three are atomic and deterministic, resolving instructions like click the submit button at runtime through AI rather than a hardcoded selector, so scripts survive redesigns. The fourth, added in version 2.0, chains them into autonomous multi step tasks when that is genuinely wanted. Version 2 also improved extraction by reading the Chrome Accessibility Tree rather than raw DOM, added self healing, and optimised model selection. It runs in TypeScript and Python, sits on top of Playwright, and works locally with any Chromium browser with no vendor account required.
Two capabilities matter for production use. Actions can be previewed before they run, so an operator sees what the AI intends rather than discovering it afterwards, and resolved actions can be cached and replayed, which converts an AI driven exploration into a repeatable deterministic workflow and cuts token spend. Browserbase describes it plainly: this is not a black box agent. An MCP integration exposes Stagehand to any external language model, and Browserbase positions Stagehand plus Claude as an alternative to OpenAI's Operator.
The commercial logic is the Red Hat model applied to browser automation: give away the framework, charge for the infrastructure. Stagehand is free forever and runs locally, but concurrent sessions, stealth mode, proxy rotation, session replay, CAPTCHA solving, Agent Identity and a Model Gateway all live on Browserbase's paid cloud. Reported adoption varies widely across sources, from 10,000 to more than 50,000 GitHub stars with around half a million weekly downloads, so treat any single figure with caution.
Vendor details
Canonical URL
https://www.stagehand.dev
Category
Browser / computer-use agent
Company status
unclear
Use cases & customers
Target customers
Deployment options
In practice
Your Playwright scripts break every time the site ships a redesign. Stagehand resolves instructions like click the submit button at runtime through AI rather than a hardcoded selector.
A fully autonomous agent is too unpredictable for production but pure code is too brittle. Stagehand lets you use natural language for unfamiliar pages and deterministic code where you already know the flow.
You want to see what the AI intends before it acts. Actions can be previewed before running and cached once resolved, turning an exploratory run into a repeatable workflow that costs fewer tokens.
Sources & related URLs
Agentic Index coverage score
10.5 / 14 capabilities · 75%
| Integrations & Tool CallingBuilt on top of Playwright so the entire Playwright ecosystem remains available, works with any Chromium browser locally, supports multiple AI model providers, integrates with Browserbase infrastructure, and exposes an MCP surface to external models. 2026-08-05 | Full |
|---|---|
| Workflow OrchestrationThe agent primitive introduced in version 2.0 chains act, extract and observe into autonomous multi step tasks, and resolved actions can be cached and replayed as repeatable workflows, so exploration converts into deterministic execution. 2026-08-05 | Full |
| Knowledge Grounding & RAGThe extract primitive and version 2's use of the Chrome Accessibility Tree give structured comprehension of page content rather than raw DOM scraping, which is grounding in the live page, but there is no retrieval corpus, index or citation layer. 2026-08-05 | Partial |
| Human Oversight & GuardrailsDistinctive and explicit: Stagehand lets you PREVIEW AI ACTIONS BEFORE RUNNING THEM, and the developer chooses which steps are natural language and which are deterministic code. Browserbase states plainly that Stagehand is not a black box agent, which is the whole design position against fully autonomous alternatives. 2026-08-05 | Full |
| Security, Identity & GovernanceAs an MIT licensed library running locally, no data need leave the customer environment at all, which is a real security property, but certifications, access controls and governance belong to whatever hosts it. Browserbase supplies Agent Identity and session controls when used on its cloud. No certifications apply to the framework itself. 2026-08-05 | Partial |
| Observability & AuditabilitySession replay and prompt observability are available when running on Browserbase, and action preview shows intent before execution, but the framework running locally provides no built in tracing or run history of its own. 2026-08-05 | Partial |
| Memory & State PersistenceAction caching persists AI resolved actions across runs so repeated workflows skip re resolution and save tokens, which is genuine cross run state, but there is no broader agent memory or accumulating task context. 2026-08-05 | Partial |
| Deployment & Data ResidencyWorks locally out of the box with any Chromium browser and no vendor account, under an MIT licence, or runs on Browserbase's managed cloud. Fully self hostable by construction, with the vendor stating explicitly that Browserbase is not required. 2026-08-05 | Full |
| Prebuilt Agents, Templates & PacksFour composable primitives and documented examples provide a starting structure, and Browserbase's Director generates Stagehand scripts from plain English as a separate product, but Stagehand itself ships no template gallery or prebuilt automation catalogue. 2026-08-05 | Partial |
| Triggers & Channel CoverageInvoked programmatically from TypeScript or Python code, or through MCP by an external model, but the framework provides no scheduling, event triggers or inbound channels of its own; those belong to whatever calls it. 2026-08-05 | Partial |
| Model Flexibility & RoutingSupports multiple AI models from launch with optimised model selection added in version 2, and on Browserbase a Model Gateway provides access to every supported model through a single API key with no separate provider accounts or credentials. 2026-08-05 | Full |
| APIs, SDKs & MCP ExtensibilityMIT licensed open source with SDKs in TypeScript and Python, built on Playwright so existing tooling composes, and an MCP integration exposing Stagehand to any external language model. 2026-08-05 | Full |
| Testing, Debugging & OptimizationAction preview supports iterative debugging, self healing was improved in version 2, and caching makes runs reproducible, but no evaluation harness, benchmark suite or regression testing framework is provided. 2026-08-05 | Partial |
| Browser & Computer UseThis is the product: natural language control of a real browser through act, extract, observe and agent primitives, clicking, typing and navigating real pages, positioned by Browserbase alongside Claude as an alternative to OpenAI's Operator. 2026-08-05 | Full |
The Agentic Index coverage score grades every vendor Full, Partial or Unable to verify against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free and open source under MIT; Browserbase infrastructure billed separately, with Director reported at $20 per month including 100 browser hours then $0.12 per hour and $12 per GB of proxy
free framework; browser hours and proxy bandwidth billed by Browserbase
What is public
The MIT licence and full source are public on GitHub. Infrastructure rates belong to Browserbase and are published on its own pricing page.
Billing mechanics
Stagehand carries no licence cost at all. Running locally, the only spend is model tokens from whichever provider the developer uses. Moving to Browserbase adds metered browser hours and proxy bandwidth, and unlocks concurrent sessions, stealth mode, session replay, CAPTCHA solving, Agent Identity and a Model Gateway that consolidates model access under one API key.
Cost watchouts
Three separate invoices are easy to miss: browser time on Browserbase, LLM tokens from your model provider, and proxy bandwidth. The framework being free does not make the workload free.
Variable cost rationale
Local use costs only model tokens, and action caching directly reduces that, but hosted use adds browser hours and proxy bandwidth metered by consumption, and an agent exploring unfamiliar pages consumes all three simultaneously.
Additional watchouts
The framework being free is not the same as the workload being cheap. Model tokens dominate cost for AI resolved actions, which is exactly why action caching exists, and teams that skip caching pay repeatedly to resolve the same selectors. Above about 500 browser hours a month, compare against self run alternatives.
Overage / add-ons
On Browserbase, browser hours beyond the plan allowance are reported at $0.12 per hour and proxy bandwidth at $12 per GB beyond the first gigabyte
Sales call required
No, self serve available
Free / trial
Completely free under the MIT licence with no usage limits, running locally against any Chromium browser without a Browserbase account
Commercial notes
This is the Red Hat model applied to browser automation, and it is worth understanding as strategy rather than pricing. Giving the framework away with genuine local capability builds adoption, and the paid infrastructure captures the moment a workload needs concurrency, stealth or proxy rotation, which is precisely when it becomes production. Independent analysis puts Browserbase's pricing as fair below roughly 500 browser hours a month and worth modelling carefully above that, with Steel positioned as the self run alternative for teams that conclude they could operate it themselves.
Key ambiguities
Stagehand itself has no price, so the meaningful figures belong to Browserbase, and the Director rates quoted here come from a competitor's review rather than a first party page. Browserbase's own tier structure was not retrieved in this pass.
Missing data
Browserbase's current tier structure and rates, and whether Model Gateway usage carries a margin over direct provider pricing.
Related vendors
- AgentQL — A self-healing query language for the web: natural language…
- Airtop — Cloud browsers for AI agents that compile plain English workflows…
- Asteroid — Supervised browser and computer use agents for regulated back…
- Astral — Browser based AI marketing agent that operates social platforms like…
- Automat AI — Managed AI agents that operate computers visually to replace legacy…
- Autotab — General AI agent that uses a mouse and keyboard like a person to do…
Alternatives to Stagehand
The closest documented capability profiles to Stagehand among browser and computer-use agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Browser Use11.0 / 14Fuller documented coverage on Observability & Auditability and Prebuilt Agents, Templates & Packs
- Emergence AI12.5 / 14Fuller documented coverage on Security, Identity & Governance and Observability & Auditability
- Browserbase12.0 / 14Fuller documented coverage on Security, Identity & Governance and Observability & Auditability
- Fellou9.0 / 14Fuller documented coverage on Memory & State Persistence
- Narada8.0 / 14A lighter documented profile than Stagehand
- Runner H8.0 / 14A lighter documented profile than Stagehand
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded