Browser infrastructure for agents: Browserbase vs Steel vs Hyperbrowser
Agent libraries like browser-use and Stagehand decide what to click; something still has to run the browsers they click in. At small scale that's Chrome on your own machine. At production scale it becomes its own operations problem: hundreds of concurrent sessions, sites that block automation, CAPTCHAs, logins that must persist between runs, and recordings you'll need when a run goes wrong. Browser infrastructure vendors sell exactly that: hosted browsers with the ugly parts handled. These three are the names buyers actually weigh (our browser-agents guide covers the other two lanes: agent libraries and MCP tool servers).
| Browserbase | Steel | Hyperbrowser | |
|---|---|---|---|
| What it sells | Managed cloud browsers with automation, search, and extraction APIs | An open-source browser API you run hosted or self-hosted | Cloud browsers with agent-focused scraping and extraction APIs |
| Open source? | No: closed infra; its open source is the Stagehand SDK and an MCP server | Yes: steel-dev/steel-browser (7.5k stars, Apache-2.0) is the actual product core, Docker-deployable | No: closed infra; peripheral MIT tooling only |
| Steward | Browserbase ($40M Series B, company blog) | Nen Labs | Hyperbrowser (Y Combinator) |
| Anti-bot and CAPTCHAs | Via partnerships (Cloudflare, Fingerprint), per its docs | Stealth mode, residential proxies, CAPTCHA solving, per its docs | "Ultra Stealth Mode"; CAPTCHA solving priced into agent steps, per its docs |
| Session recording | Yes, every agent step plus live view | Yes, session replays | Yes, video recordings |
| Logged-in state between runs | Context-style persistence | Persistent profiles | Profiles via API |
| Pricing shape (checked 2026-08-12) | Base fee plus usage: free tier, then $20 and $99/month tiers with browser-hour overage around $0.10 to $0.12/hr | Usage-based with plan fees: $0/month plus $30 one-time credit to start, $250/month at scale | Pure credits (1,000 = $1): $0.10 per browser-hour, $0.02 per agent step |
| Works with | Stagehand natively; LangChain, CrewAI, Mastra | Framework-agnostic: any Playwright or Puppeteer stack | Claude, OpenAI, Gemini, and browser-use agents built in |
Only Steel has an open-source core, so this page carries no star-ranked table; Steel's list entry lives in Libraries and SDKs. Pricing shapes summarize each vendor's own pricing page on the date shown; check them before committing.
Pick by situation
Three independent comparison write-ups (APIScout, PkgPulse, and Steel's own head-to-head, bias noted) land on the same split, which is rare enough to trust:
- You want the safest managed default with the most polish → Browserbase. The most mature tooling story (Stagehand is its SDK), session visibility on every agent step, and the most generous named free tier of the three. The trade: fully closed, and features track plan level.
- Openness or self-hosting is a buying criterion → Steel. The only one whose core you can read, run in your own Docker, and walk away from without losing the integration; the hosted product is the same code with the ops handled. This is also the pick when procurement rules out closed infra handling logged-in sessions.
- You're scraping or crawling at volume, agent-first → Hyperbrowser. Credit pricing that meters agent steps rather than plans, stealth as the headline feature, and built-in support for driving it from browser-use and the major model providers.
Also in the lane, one tier out: Browserless, the long-running self-hosted headless-Chrome standard, and Anchor Browser, a newer managed entrant focused on agent identity. Neither shows up in buyer comparisons as often as these three yet.
The question to ask first
Does your agent hold credentials? A hosted browser that logs into your users' accounts concentrates real risk with the vendor: recordings, cookies, and persistent profiles all live on someone else's infrastructure. If the answer is yes and that sentence made you uncomfortable, Steel's self-hosted mode is the escape hatch the other two don't offer. If the answer is no (public-web scraping, testing, research), pick on developer experience and unit economics, which is the fight Browserbase and Hyperbrowser are actually having.
Part of best-of-Agent-Harnesses. New to this decision? Start with How to pick a harness. Spot an error or a stale claim? Open an issue.
Agents can query this page's data directly: claude mcp add agent-harnesses -- uvx agent-harnesses-mcp (MCP server).