Six platforms that let an agent read, click and log into real pages — ordered by how much of the job each one takes off you, with the work it is genuinely best at and where it stops.
There is no single ranking here, because the right answer depends on whether your agent reads pages or acts on them. So this is ordered from the widest surface — a platform that will also do search, identity and model routing — down to the narrowest single endpoint. Wayfern sits where it honestly belongs on that axis, not at the top.
A managed cloud browser fleet that lists six platform components — Browsers, Search & Fetch, Agent Identity, Functions, Model Gateway and Stagehand — and is driven from Playwright, Puppeteer or Selenium.
Best for Agents that need a whole platform under one vendor, and teams porting existing Selenium code: the docs publish a Selenium quickstart covering both Node.js and Python alongside the Playwright, Puppeteer and Stagehand ones.
A proxy platform — datacenter, ISP, residential and mobile networks — with managed unblocking APIs, a hosted Browser API, a SERP API, a dataset marketplace and an MCP server on top.
Best for Agents whose hard problem is the exit IP. Residential targeting goes down to country, US state, city, US ZIP and ASN, and SERP and Web Unlocker bill per 1,000 successful requests rather than per attempt.
An open-source context API with scrape, crawl, map, search, interact, agent and batch-scrape endpoints that return markdown, cleaned HTML, screenshots or schema-driven JSON.
Best for RAG and research pipelines that want one call per page, explicit cache controls (maxAge, minAge, storeInCache), and the option to run the engine in-house — though the self-hosted build does not include Fire-engine.
Cloud browsers reachable three ways: a WebSocket for existing Puppeteer or Playwright code, stateless REST endpoints, and BrowserQL, a GraphQL API described as stealth-first.
Best for Work that needs a specific engine or a specific challenge solved. It documents six browsers — Chromium, Chrome, Stealth, Firefox, WebKit and Edge — and a captcha enum spanning Cloudflare, DataDome, PerimeterX, Akamai and AWS WAF.
Anti-detect browser sessions over CDP plus a Web Data API of eight metered endpoints and live search, exposed to models by an eleven-tool MCP server, all drawing on one credit balance.
Best for Agents that have to stay logged in and keep one exit identity for a whole run: the session is the CDP socket you hold open, and the residential exit is pinned by country, region, city, ISP and sticky session id on the connect URL.
A single GET endpoint that fetches one URL with a headless browser and returns HTML, or — via return_page_markdown, return_page_text, screenshot, extract_rules or a natural-language ai_query — markdown, page text, an image or just the fields you asked for.
Best for Cost-modelled one-shot fetches. The credit table is published per configuration (1 / 5 / 10 / 25 / 75), Auto-Mode escalates from the cheapest configuration that works, and a request where every configuration fails costs 0 credits.
A request/response API sends one call per page and throws the browser away afterwards. A held CDP session keeps one browser, one cookie jar and one exit identity alive for as long as your agent needs it.
Almost every disappointment with agent web tooling traces back to this fork. If the agent only needs to read — a documentation page, a pricing page, a search result — the request shape is strictly better: it is cheaper per page, it parallelises trivially, and Firecrawl or ScrapingBee will hand back markdown you can put straight in a prompt. Firecrawl even lets you trade freshness for spend explicitly with maxAge, minAge and storeInCache.
The moment the agent has to act — log in, fill a multi-step form, dismiss an interstitial, keep a cart alive across ten navigations — a per-request API starts fighting you, because the thing you need to persist is exactly the thing it discards. That is where the session-shaped tools live: Browserbase, Browserless and Wayfern all hand you a real browser you drive yourself.
Session-shaped is not automatically freer, though. Bright Data’s hosted Browser API confines a session to one domain, times out after five idle minutes, and caps total session duration at 60 minutes — fine for a scrape, awkward for a long agent run. On Wayfern the session simply is the open CDP socket: there is no create-session call, and it is billed at 8 credits per open-page hour with a 60-second minimum on every page, so short-lived pages are not free.
Browserbase is the tool on this list aimed squarely at the same job, so it is the honest comparison — and it wins several of these rows. Every Browserbase cell was read from their own published documentation; anything their docs do not state is marked rather than guessed.
| Capability | Wayfern | Them |
|---|---|---|
| Client protocol | CDP over a WebSocket. Playwright’s connectOverCDP and Puppeteer’s connect both attach; there is no WebDriver path. | Playwright, Puppeteer or Selenium, with a published quickstart for each and Selenium documented from both Node.js and Python. |
| Starting a session | A session exists because a CDP socket is open — no create-session call. Each connection gets its own browser context. | A Create Session API taking projectId, browserSettings, timeout, keepAlive, proxies, proxySettings, region and userMetadata. |
| State that outlives a session | Not offered. There is no stored-context API, so cookies and localStorage have to be replayed by your own code. | Contexts persist cookies, localStorage, IndexedDB, session storage, service workers and form autofill between sessions — though not the HTTP cache. |
| Session recording and replay | Not offered. The gateway neither records nor replays a session. | recordSession and logSession are settings on the Create Session API. |
| Exit geo control | country, region, city, isp and a sticky session id on the connect URL, from a live catalog you can browse before connecting. | A required country plus optional state and city over built-in proxies stated to cover 201 countries; proxies default to off, and proxies: true is a best-effort US exit that may route via a nearby country. |
| What bandwidth costs | Datacenter egress is free; residential is metered per byte at 320 credits per GB, sampled every five seconds, and a session that can no longer pay is closed. | Proxy GB is a metered line on the published plans: 1GB then $12/GB on Developer, 5GB then $10/GB on Startup. |
| Captcha handling | An explicit solve call covering reCAPTCHA v2, v3 and Enterprise, hCaptcha, Turnstile, FunCaptcha, GeeTest and image-to-text, charged only on success. | On by default — solveCaptchas “Defaults to true” — with custom captcha image and input selectors available. |
| Strongest anti-detect tier | The same anti-detect browser serves every plan; there is no higher stealth tier to upgrade into. | Verified sessions run on a purpose-built Chromium with real browser fingerprints, but the docs state Verified is offered on the Scale plan, which is custom-priced. |
| Concurrency ceiling | 20 concurrent pages on Developer, 100 on Business, 400 on Scale; each link token also carries its own cap, and an over-cap connection is refused at the handshake. | Free 3 concurrent and 5 sessions per minute, Developer 25 / 25, Startup 100 / 50, Scale 250+ / 150+; over-limit requests return HTTP 429 and the docs say the request “was effectively dropped”. |
| Agent-facing tool surface | A remote Streamable HTTP MCP server exposing eleven tools — web search, search-engine listing and nine Web Data tools — on a token carrying the mcp scope. | Six platform components are listed: Browsers, Search & Fetch, Agent Identity, Functions, Model Gateway and Stagehand. |
| Free tier | None. The cheapest way in is Developer at $19/month with 1,520 credits. | Free at $0/month: 3 concurrent browsers, 1 browser hour, 3 agent runs, 15-minute sessions, no captcha solving and no stealth mode. |
| Self-hosting | Not offered as a customer-deployable product. | No self-hosted or on-premise deployment is documented; “self-hosted” in their docs means running your own Chromium instead. |
Headline prices are close to useless here, because these products meter completely different things. What matters is which of your behaviours the meter taxes.
Per browser-time (Browserbase, Browserless). You pay for wall-clock time a browser is open, so slow pages, sleeps and forgotten sessions are the expensive mistakes. Browserbase publishes browser hours per plan with per-hour overage; Browserless sells units, and a unit is a block of browser time of up to 30 seconds per browser connection, with per-unit overage. Both also cap how long one session may run — Browserless’ published max session length runs from 1 minute on Free to 60 minutes on Scale.
Per page (Firecrawl). Scrape, crawl and map each cost 1 credit per page, which makes cost modelling for a crawl trivially predictable. The tax lands on breadth: a 50,000-page crawl is 50,000 credits whether or not you needed all of it, and PDFs bill per page.
Per request, tiered by difficulty (ScrapingBee). 1 credit for a plain fetch, 5 with JavaScript rendering, 10 or 25 for premium proxies, 75 for the stealth tier — so the meter taxes how hard the target is, not how long you took. Auto-Mode makes that explicit by climbing from the cheapest configuration and billing only the one that worked.
Per successful request, plus bandwidth (Bright Data). SERP API and Web Unlocker are priced per 1,000 successful requests and do not bill failures by default, which pushes unblocking risk onto the vendor — but enabling custom headers or cookies flips both to billing 100% of requests, failures included.
One credit balance across everything (Wayfern). Sessions burn 8 credits per open-page hour with a 60-second floor per page; the Web Data endpoints meter per unit — scrape 1 credit per page, crawl 1 credit per successful page, extract 1 credit per source page, map 2 credits per 10 URLs, screenshot 5 credits per capture — with design intelligence from 5 credits a site for fonts up to 10 for a brand profile or styleguide; residential egress is 320 credits per GB and datacenter egress is free; asking a scraped page for a screenshot, PDF or JSON output adds +4 once, and an LLM summary +2. The tax is on holding browsers open when a single fetch would have done — and the honest catch is that there is no free tier at all: Developer is $19/month for 1,520 credits, Business $99 for 9,000, Scale $499 for 46,000.
Four common agent workloads, and the tool on this list each one actually points at.
Your agent reads and summarises. It needs clean markdown from public pages and nothing else. Firecrawl or ScrapingBee: one call per page, LLM-ready output, no session to manage, and caching or Auto-Mode to keep the bill down. Reach for a browser platform here and you are paying for wall-clock time you never use.
Your agent acts. It logs in, fills forms, clicks through flows, and needs the cookie jar and the exit IP to stay put across many navigations. That is a held browser: Browserbase, Browserless or Wayfern. Prefer Browserbase if you want session recording and saved contexts out of the box, Browserless if you need a non-Chromium engine, and Wayfern if the run has to keep one residential identity pinned to a country, city and ISP from the first request to the last.
Your blocker is the IP, not the browser. The pages render fine, you just get blocked or geo-fenced. Bright Data, whose residential targeting reaches US state, city, ZIP and ASN across four networks — with the caveat that residential access is not self-serve: personal-email accounts are ineligible, approval is human-reviewed KYC, and traffic outside your approved use case is refused.
You must run it on your own infrastructure. On this list only Firecrawl publishes a self-hosting guide, and it is honest about the limits: self-hosted instances do not get Fire-engine, so the advanced IP-block and bot-detection handling stays behind the hosted service. Browserbase documents no on-premise deployment, and Wayfern does not ship one either.
Mostly, yes. Firecrawl publishes an official MCP server so agents in Cursor, Claude, Windsurf and other MCP-compatible clients can search and scrape directly, and Bright Data’s documentation lists an MCP server as one of its four platform areas, beside proxy infrastructure, web access APIs and data feeds. Wayfern exposes a remote Streamable HTTP MCP server with eleven tools — web search, engine listing, and the nine Web Data tools — on an API token carrying the mcp scope, and every fetching tool states its credit price in its own description so the model can prefer the cheapest one that answers the question.
Any of the session-shaped ones, because you drive the login yourself: Browserbase, Browserless and Wayfern all give you a real browser over CDP. The difference is what happens afterwards. Browserbase can persist that logged-in state into a Context and restore it in a later session. Wayfern has no stored-context API, so you keep the socket open for the whole run — or capture and replay the cookies yourself — while holding one sticky residential exit so the session does not change IP mid-flow.
Usually not for public documentation, blogs or product pages — datacenter egress is free on Wayfern and the default nearly everywhere else. You need residential when the target scores the IP: retail, travel, social and anything behind a serious bot wall. Budget for it, because it is the line item that moves: Wayfern meters residential per byte at 320 credits per GB and closes a session that can no longer pay, and Browserbase bills proxy GB on top of browser hours. Note also that Browserbase’s built-in residential proxies come from third-party providers that restrict banking, government, streaming, ticketing, webmail and gambling domains.
The two that document the behaviour reject rather than queue. Browserbase returns HTTP 429 and its docs state the request “was effectively dropped”. Firecrawl’s rate limits are per team, shared by every API key on it, and also return 429. Wayfern refuses the connection at the WebSocket handshake with an explicit close code, both against the plan cap — 20 concurrent pages on Developer, 100 on Business, 400 on Scale — and against the per-token cap you set on each link. Either way, backpressure and retries are your code’s job.
Mostly yes, and this is where Wayfern is the weakest of the six. Browserbase, Firecrawl and Browserless all publish a $0 tier; ScrapingBee gives 1,000 API credits with no credit card; Bright Data publishes 5,000 free monthly credits for its Web Unlocker, SERP, Web Scraper and Scraper Studio products. Wayfern has no free tier — platform access needs an active paid plan or a remaining credit-pack balance, starting at Developer at $19/month. What it does not do is gate features by plan: the anti-detect browser, residential targeting, captcha solving, MCP and every Web Data endpoint are on the cheapest plan, and the tiers differ only in credits, concurrency and seats.
Three places, plainly. There is no free tier. There is no Selenium or WebDriver support at all — Playwright and Puppeteer over CDP only — so Selenium codebases point at Browserbase, Browserless or Bright Data instead. And there is no session recording and no saved browser contexts, both of which Browserbase documents and some teams will consider non-negotiable for debugging an agent that misbehaves in production.
Browser sessions, the Web Data API, search and captcha solving all draw on one credit balance. Start on any of them.