---
type: analysis
slug: guide-browser-agents-stack
title: "Browser automation for agents: Browserbase, Stagehand, Playwright MCP"
description: "Browser automation for AI agents splits into two layers: where the browser runs and how the model drives it. Browserbase is a hosted platform that gives agents fleets of headless cloud browsers plus Fetch, Search, Identity (stealth, CAPTCHA solving, proxies, authentication) and session observability"
published_at: 2026-11-03T09:00:00.000Z
topics: ["tools","developer_tools"]
source_urls: ["https://docs.stagehand.dev/v3/first-steps/introduction","https://docs.stagehand.dev/v3/best-practices/caching","https://github.com/microsoft/playwright-mcp","https://www.browserbase.com/pricing","https://steel.dev/pricing","https://browser-use.com/pricing","https://github.com/browserbase/stagehand","https://github.com/browserbase/stagehand/releases"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
Browser automation for AI agents splits into two layers: where the browser runs and how the model drives it. Browserbase is a hosted platform that gives agents fleets of headless cloud browsers plus Fetch, Search, Identity (stealth, CAPTCHA solving, proxies, authentication) and session observability, billed per browser hour. Stagehand is Browserbase's open-source MIT SDK that mixes natural-language act/extract/observe/agent calls with deterministic Playwright-style code and can run on a local browser or on Browserbase. Playwright MCP is Microsoft's MCP server that lets any MCP client (Claude Code, Codex, Cursor, VS Code and others) control a local browser through accessibility snapshots rather than screenshots. It is for engineers building agents that log in, fill forms or extract data from sites without APIs.

*Field guide compiled 2026-10-23 from the official documentation, pricing pages and changelogs linked below; every line links to the page that states it.*

## What it does

- **Act.** Execute browser actions from a natural-language instruction; Stagehand says act() self-heals when a site redesigns its form. How: await stagehand.act("click the login button") [source](https://docs.stagehand.dev/v3/first-steps/introduction)
- **Extract.** Pull schema-validated structured data from a page. How: await stagehand.extract("extract the price", z.number()) [source](https://docs.stagehand.dev/v3/first-steps/introduction)
- **Observe.** Discover available actions on a page and return real selectors, so code can fill credentials without passing them to the model. How: await stagehand.observe("find submit buttons") [source](https://docs.stagehand.dev/v3/first-steps/introduction)
- **Agent.** Automate whole multi-step workflows autonomously, including computer-use (CUA) mode with a chosen model. How: stagehand.agent({ mode: "cua", model: "..." }).execute("apply for this job") [source](https://docs.stagehand.dev/v3/first-steps/introduction)
- **Caching Actions.** Two caching strategies, Browserbase Cache (managed, server-side) and Local Cache, to reduce LLM cost and speed up repeated automations. How: See Stagehand Best Practices Caching Actions [source](https://docs.stagehand.dev/v3/best-practices/caching)
- **Playwright MCP.** MCP server that drives a browser via Playwright's accessibility tree, so no vision model is needed and tool application is deterministic. How: Add {"command": "npx", "args": "@playwright/mcp@latest"} to the MCP client config; e.g. claude mcp add playwright npx @playwright/mcp@latest or codex mcp add playwright npx "@playwright/mcp@latest" [source](https://github.com/microsoft/playwright-mcp)
- **--caps.** Opt-in extra Playwright MCP capabilities such as vision (coordinate-based), pdf, devtools, network, storage and testing tools. How: npx @playwright/mcp@latest --caps=vision,pdf (env PLAYWRIGHTMCPCAPS) [source](https://github.com/microsoft/playwright-mcp)
- **--extension.** Connect Playwright MCP to a running Edge/Chrome instance and reuse your logged-in tabs and browser state. How: Install the Playwright Extension, then run with --extension [source](https://github.com/microsoft/playwright-mcp)

## Modes

- **Persistent profile.** Playwright MCP default: logged-in state is stored in a per-workspace profile under ms-playwright/mcp-{channel}-{workspace-hash}; override with --user-data-dir. [source](https://github.com/microsoft/playwright-mcp)
- **Isolated.** Playwright MCP --isolated: each session uses a throwaway profile; seed login state with --storage-state. [source](https://github.com/microsoft/playwright-mcp)
- **Browser Extension.** Playwright MCP attaches to your existing Chrome/Edge tabs and sessions. [source](https://github.com/microsoft/playwright-mcp)
- **Playwright CLI.** Microsoft recommends CLI + SKILLs over MCP for coding agents because it avoids loading large tool schemas and accessibility trees into context. [source](https://github.com/microsoft/playwright-mcp)
- **cua.** Stagehand agent mode that runs a computer-use model to drive the browser. [source](https://docs.stagehand.dev/v3/first-steps/introduction)
- **localBrowser.** Stagehand on a local browser with userDataDir so cookies persist between runs, instead of a hosted Browserbase session. [source](https://github.com/browserbase/stagehand)

## Pricing and limits

- **Browserbase Free Plan:** $0/mo. 3 concurrent browsers, 1 browser hour, 3 Agent runs, 1,000 Search and 1,000 Fetch calls, 15 minutes/session, $5 in model tokens, no CAPTCHA solving, 7-day retention [source](https://www.browserbase.com/pricing)
- **Browserbase Developer Plan:** $20/mo. 25 concurrent browsers, 100 browser hours then $0.12/browser hr, 15 Agent runs, 1 GB proxies then $12/GB, auto CAPTCHA solving, basic stealth [source](https://www.browserbase.com/pricing)
- **Browserbase Startup Plan:** $99/mo. 100 concurrent browsers, 500 browser hours then $0.10/browser hr, 50 Agent runs, 10,000 Fetch calls then $0.5/1k, 5 GB proxies then $10/GB [source](https://www.browserbase.com/pricing)
- **Steel Launch:** $0 per month + usage. $30 one-time credits, 10 concurrent sessions, sessions up to 15 minutes; browser hours $0.10, CAPTCHA solves $3/1k, proxy $10/GB [source](https://steel.dev/pricing)
- **Steel Scale:** $250 per month + usage. $100 credits/month, 100 concurrent sessions, sessions up to 1 hour; browser hours $0.08, CAPTCHA solves $1/1k, proxy $6/GB [source](https://steel.dev/pricing)
- **Browser Use cloud browsers:** $0.02 / hour. Pay as you go, plus $5 / GB residential proxy; browser agents billed at model + 20% [source](https://browser-use.com/pricing)

## Recent changes

- 2026-08-28: Stagehand server v3.7.6 and related Stagehand releases published on GitHub. [source](https://github.com/browserbase/stagehand/releases)

## Alternatives

- **Steel.** Hosted browser infra priced as plan plus usage: $0.10 browser hour on Launch and $0.08 on Scale ($250/month), with CAPTCHA solves billed per 1k; stealth browser only on Enterprise. [source](https://steel.dev/pricing)
- **Browser Use.** Pay-as-you-go cloud browsers at $0.02/hour with no subscription, and hosted browser agents billed at model cost + 20%. [source](https://browser-use.com/pricing)
- **Playwright CLI.** Same Playwright engine exposed as CLI commands and skills, more token-efficient than the MCP server for coding agents. [source](https://github.com/microsoft/playwright-mcp)

## Traps

- A Playwright MCP persistent profile can be used by only one browser instance at a time, so concurrent MCP clients in the same workspace conflict; use --isolated or a distinct --user-data-dir per client. [source](https://github.com/microsoft/playwright-mcp)
- In Playwright MCP isolated mode, closing the browser discards all storage state for the session, so logins are lost unless you pass --storage-state. [source](https://github.com/microsoft/playwright-mcp)
- Browserbase Free caps sessions at 15 minutes, includes 1 browser hour and has no CAPTCHA solving; CAPTCHA solving starts on Developer. [source](https://www.browserbase.com/pricing)
- MCP browser tools load large tool schemas and verbose accessibility trees into the model context, which Microsoft flags as a token cost for coding agents. [source](https://github.com/microsoft/playwright-mcp)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Tools](https://newruntime.com/topics/tools.md): Explore the tools topic hub.
- topic hub: [Developer Tools](https://newruntime.com/topics/developer-tools.md): Explore the developer_tools topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)