Browser Use To Stagehand

browser-use → Stagehand on Browserbase (/browser-use-to-stagehand)

Convert a browser-use (Python) script into an idiomatic Stagehand v3 (TypeScript) script on Browserbase, choosing the right level of determinism at each step rather than producing a one-to-one agentic copy.

Core principle: browser-use is agentic-by-default (the LLM decides every action). Stagehand lets you choose how much AI to use. A good migration replaces opaque agent loops with an inspectable, mostly-deterministic pipeline — using AI only where the page is genuinely unpredictable. This is a refactor with judgment, not a transpile.

Source of truth & versions. This skill’s durable value is the judgment — the determinism spectrum and the decompose-vs-agent decision — not the API specifics, which drift every release. The code mappings here are a snapshot validated against @browserbasehq/stagehand 3.6.x and browser-use 0.13.x (2026-06). On any conflict, the live docs win — always verify against the installed package and these sources before emitting code:

• Stagehand v3: https://docs.stagehand.dev/v3 (opens in a new tab) · installed types: node_modules/@browserbasehq/stagehand
• Browserbase: https://docs.browserbase.com (opens in a new tab)
• browser-use: https://docs.browser-use.com (opens in a new tab)

If the installed Stagehand major is not 3, treat this skill as conceptual only and follow the live docs for every signature.

Reference files (read as needed)

• references/api-mapping.md (opens in a new tab) — the mechanical browser-use → Stagehand mapping: variant detection, the full feature table, before/after code, Browserbase platform options, and v3 version gotchas. Read this for any non-trivial construct.
• references/determinism.md (opens in a new tab) — how to choose agent() vs act/extract/observe vs cached observe→act. The decision tree. Read this when deciding how to translate an Agent(task=…).
• references/trace-assisted.md (opens in a new tab) — the optional “run it on Browserbase, read the logs, then rewrite” workflow for opaque/flaky scripts.
• — the human migration guide: philosophy shift, feature mapping, the determinism spectrum, and a recommended migration path.

Workflow

1. Get the source

Obtain the browser-use script(s). If the user only described a script, ask for the file(s). Note the target: TypeScript Stagehand on Browserbase unless they say otherwise.

First, gate on scope — is this even migratable? Not every browser-use file is an Agent(task=…) script. If the source is browser-use running as an MCP server (uvx browser-use --mcp, a mcpServers config) there is no Stagehand equivalent — flag it as out of scope, don’t invent one (see api-mapping §3.7b). If the browser-use call is embedded in a larger app (a class/tool wrapper, web route, queue task), convert only the browser-use surface and preserve the surrounding app glue — see api-mapping §3.8.

2. Detect the browser-use variant

Identify legacy (pre-0.12) vs stable vs Rust beta (only when imports come from browser_use.beta) — see api-mapping §1. Note: the classic top-level from browser_use import Agent, ChatBrowserUse surface is alive and well in 0.13.x — ChatBrowserUse alone is not a beta tell; only a browser_use.beta import is. All variants translate identically, so when unsure, proceed with the stable mapping. Normalize legacy names before translating. State which variant you found.

3. Inventory the script

Extract a structured inventory before writing any TypeScript:

• Task(s) — the task= string(s); split each into its implied ordered steps.
• Model — the Chat* provider + model id.
• Browser config — local vs cdp_url/Browserbase; headless; proxies; user_data_dir/storage_state.
• Structured output — any output_model_schema Pydantic models.
• Secrets — sensitive_data, env-var usage, login flows.
• Guardrails — allowed_domains, max_steps.
• Custom actions — @tools.action / Controller functions, and whether each is a deterministic side-effect or an agent capability.
• Setup — initial_actions, secondary models (page_extraction_llm, planner_llm).

4. Decide the determinism level per step

For each step from the inventory, apply the decision tree in determinism.md:

• Navigate to a known URL → page.goto(url) on the Stagehand page (no AI).
• On-page action → act("…"); if it repeats, observe() once then replay act(action) (no LLM call).
• Reading data → extract("…", zodSchema).
• Genuinely open-ended → keep stagehand.agent().execute(...) (tightened with maxSteps/systemPrompt).

Default to decomposition when the flow is known; keep agent() only where it isn’t. For a first lift-and-shift, a faithful agent() translation is acceptable — say so and note the optimization path.

5. Produce the Stagehand v3 rewrite

First, verify the API. Before writing, confirm the exact signatures you’re about to use against the installed package (node_modules/@browserbasehq/stagehand types) or https://docs.stagehand.dev/v3 (opens in a new tab). The mappings below are a 3.6.x snapshot; if anything differs in the installed version, the installed version wins. Then emit runnable TypeScript. Always:

• import { Stagehand } from "@browserbasehq/stagehand"; and import { z } from "zod"; when extracting.
• Get the page via const page = stagehand.context.pages()[0];.
• Call AI methods on the instance: stagehand.act(...), stagehand.extract(...), stagehand.observe(...) — never page.act(...).
• Set the model as a "provider/model" string.
• Default to env: "BROWSERBASE"; show env: "LOCAL" as the dev option.
• Pass secrets via variables and process.env, never hardcoded.
• await stagehand.init() at the start, await stagehand.close() in a finally.

Include the project setup so it runs (see the templates below).

6. Write the migration summary

Alongside the code, produce a short summary:

• Variant detected and the determinism choices made (which steps became deterministic vs AI vs agent), with the reasoning.
• Needs human review — anything that didn’t map 1:1: lost allowed_domains guardrails, custom-action logic, secondary-model intent, ambiguous task strings.
• Recommended next step — Browserbase Context for auth reuse, caching for production, or the trace-assisted path if the flow was opaque.

7. Offer the trace-assisted path (only if warranted)

If the source was one large opaque agent(task=…), was flaky, or your rewrite can’t be confidently mapped, offer the trace-assisted workflow (trace-assisted.md): run the original on Browserbase, pull sessions.logs.list, and rewrite from observed behavior. Don’t run anything without the user’s go-ahead.

Output templates

package.json

{
  "name": "stagehand-migration",
  "type": "module",
  "scripts": { "start": "tsx index.ts" },
  "dependencies": {
    "@browserbasehq/stagehand": "^3.0.0",
    "dotenv": "^16.0.0",
    "zod": "^3.25.0"
  },
  "devDependencies": { "tsx": 

Add "ai": "^5.0.0" (Vercel AI SDK) only if a custom browser-use action maps to an agent tool. Pin v5, not v4 — Stagehand 3.6.x bundles ai v5 and types agent({ tools }) as the v5 ToolSet, where a tool’s schema field is inputSchema. The v4 tool() helper emits parameters instead and will fail to type-check against Stagehand’s v5 ToolSet. If you can’t control the hoisted ai version, skip the tool() helper and pass a plain object { description, inputSchema: zodSchema, execute } — it satisfies the v5 ToolSet regardless of which ai major resolves.

.env

BROWSERBASE_API_KEY=...
BROWSERBASE_PROJECT_ID=...
ANTHROPIC_API_KEY=...   # or the provider matching your model string

index.ts skeleton (decomposed, the preferred shape)

import "dotenv/config";
import { Stagehand } from "@browserbasehq/stagehand";
import { z } from "zod";

async function main() {
  const stagehand = new Stagehand({
    env: "BROWSERBASE",
    model: "anthropic/claude-sonnet-4-6"

Validation checklist (before declaring done)

• Not done: AI methods are on the instance (stagehand.act/extract/observe), not the page.
• Not done: Page obtained via stagehand.context.pages()[0].
• Not done: Model is a "provider/model" string; the matching provider key is in .env.
• Not done: extract uses a zod schema; zod is in dependencies.
• Not done: Secrets use variables + process.env; nothing hardcoded.
• Not done: init() / close() present; close() in finally.
• Not done: Each browser-use step is accounted for, placed deliberately on the determinism spectrum.
• Not done: Migration summary lists determinism choices and “needs human review” items.

Common mistakes to avoid

• Copying v2 syntax (page.act(), stagehand.page, modelName/modelClientOptions, enableCaching) from old blog posts. Use v3 — see api-mapping “Version notes”.
• Translating every step into act() — navigate with page.goto and cache repeatable steps via observe→act; don’t spend an LLM call on every action.
• Defaulting everything to agent() — that just reproduces browser-use’s non-determinism in a new framework. Decompose where the flow is known.
• Silently dropping allowed_domains — Stagehand has no domain firewall; flag it for review.
• Inventing Browserbase/Stagehand options — if unsure of a field, check https://docs.stagehand.dev/v3 (opens in a new tab) / https://docs.browserbase.com (opens in a new tab) rather than guessing.