Compare / other services
cdpfleet vs Steel
The open-source browser API for agents against a browser fleet: per-minute Chromium with proxies, captcha solves and a 15-minute-to-24-hour session cap by plan, or twelve real browsers metered per second with your own proxies. Prices, limits and features side by side.
Steel is the open-source browser API built for AI agents: a Chromium sandbox you reach over CDP with Puppeteer, Playwright or Selenium, with rotating proxies, captcha solving, stealth configs, live session viewers and replays, Credentials for sign-ins, Browser Tools (scrape, screenshot and PDF in one call), a hosted MCP server and a cookbook of agent integrations. The core (steel-browser, Apache 2.0) can be self-hosted; Steel Cloud is the hosted version, billed by the minute. cdpfleet is a fleet of twelve real desktop browsers — Google Chrome, Microsoft Edge, Brave, Opera, Naver Whale, Yandex Browser, Chromium, Patchright, Rebrowser, Camoufox, Firefox and WebKit — launched with one HTTPS call and driven over the Playwright protocol from five languages, through proxies you bring. Both are "a browser in the cloud by the hour"; they differ on what the browser is, how long it may live, and what else comes in the box — including Steel's Browser Tools next to cdpfleet Functions, jobs and monitors.
| cdpfleet | Steel | |
|---|---|---|
| Browsers | 12 — Google Chrome, Microsoft Edge, Brave, Opera, Naver Whale, Yandex Browser, Chromium, Patchright, Rebrowser, Camoufox, Firefox, WebKit. Chrome with stable, beta and dev channels and pinned previous majors; Camoufox with pinned versions. | Chromium. A "Stealth Browser" tier is an Enterprise feature. |
| How you connect | Playwright protocol: chromium/firefox/webkit.connect(wsUrl) with the matching Playwright client (1.60; 1.52 for Rebrowser) — Node.js, Python, Java, C# and Go. Or "cdp": true on the Chromium-based browsers for a cdpUrl: Puppeteer, connectOverCDP or any CDP client (CDP & Puppeteer). | CDP WebSocket: Puppeteer, Playwright (Node and Python) via connectOverCDP, Selenium; SDKs for TypeScript, Python, Go and Rust; REST Browser Tools for scrape, screenshot and PDF (below). |
| Session length | Up to 24 h per session on every plan (shared threads: also capped by your remaining daily minutes). Idle timeout 120 s by default, up to 10 min on shared threads and 24 h on dedicated. | 5 minutes by default; the hard cap is set at creation and can't be extended: 15 minutes (Launch), 1 hour (Scale), up to 24 hours (Enterprise). Optional inactivityTimeout releases idle sessions early. |
| Metering | Per second, as thread-seconds: headless browser = 1 thread, headful = 2. | Per minute, rounded up: $0.10/browser-hour (Launch), $0.08 (Scale), custom (Enterprise). |
| Plans | Free Community plan: 2 threads, 60 browser-minutes per thread a day, 15-minute sessions, proxies included. Paid from $7.50/month (5 shared threads, 60–480 browser-minutes a day per thread); dedicated threads for 24/7 use at $8.50/thread/month. Pricing. | Launch: $0/month + usage, $30 one-time credits (valid 90 days), 10 concurrent, 60 API requests a minute. Scale: $250/month + usage, $100 credits/month, 100 concurrent, 600 requests a minute. Enterprise: custom, 1,000+ concurrent. On Launch, Steel's proxies and CAPTCHA solving need $10 of paid balance first (pricing-and-limits page; the proxy and CAPTCHA pages only say verification is required). |
| One browser-hour | Shared: $0.02–$0.05 at full use; dedicated $0.01. A browser running 24/7 is one dedicated thread, $8.50/month. | $0.10 (Launch) or $0.08 (Scale) plus the plan fee. A browser running 24/7 would need a new session every 15–60 minutes and cost ≈ $58–$72/month in hours alone (Scale: + $250 plan fee). |
| Concurrency | The threads you buy: 5 to 10,000 (more on request); each session takes 1–4 threads by type. | 10 (Launch), 100 (Scale), 1,000+ (Enterprise); reserved browser pools on Enterprise. |
| Proxies | Bring your own HTTP, HTTPS or SOCKS5 proxies — single, pool, or per-host rules (proxy_rules) — swap a running session's proxy live, see traffic per destination host and per rule; traffic through them is free. Or the cdpfleet residential proxy: 4 tiers from $0.50 to $5/GB from one prepaid balance ($5 free credit), mixable per host — e.g. Tier 4 for CDN assets, Tier 1 for pages. A launch without a proxy uses the free Community proxy — best effort, no guarantees. | Built-in rotating residential proxies at $10/GB (Launch) or $6/GB (Scale), 200+ countries with state and city targeting; your own proxy (proxyUrl) with no Steel bandwidth charge; dedicated IPs $5/month each on Scale and Enterprise. |
| Stealth | Real browsers by default (genuine Chrome, Edge, Brave…), headful by default, and three stealth builds — Patchright, Rebrowser, Camoufox — on every plan. See headless tells. | Stealth configs on all plans; the dedicated Stealth Browser only on Enterprise. |
| Captcha solving | Not built in. Use any solver through Playwright. | Built in: $3 per 1,000 solves (Launch), $1 per 1,000 (Scale). reCAPTCHA v2/v3, Turnstile, image and slider CAPTCHAs are solved; DataDome, Imperva, AWS WAF and FunCAPTCHA are detected but not solved. |
| Watching a session | Live view with human takeover — Watch / Take control in the dashboard, or a link from the API — and Launch interactive: open any browser on a start page straight into live view from the dashboard's code generator (start_url over the API). Session recordings as MP4 (480p free, 720p or 1080p on a plan), and JPEG frames to your code with page.screencast. | Live view (WebRTC, embeddable), every session recorded automatically with MP4/HLS replays, and agent traces synced to the video; data retained 7 days (Launch) or 14 days (Scale). |
| Persistence and sign-in | Persistent profiles keep cookies, local storage and IndexedDB between sessions (paid plans); Playwright storageState; keep_alive to reconnect to the same browser on dedicated threads. No credential vault. | Session contexts reusable across runs; Profiles that keep the whole user-data directory (300 MB, deleted after 30 days unused); Credentials API for stored logins and automatic sign-in (beta, free during the beta, not in self-hosted Steel). |
| Scraping in one call | Functions: screenshot, PDF, Markdown, scrape (HTML, Markdown, text, links, screenshot, XHR calls, cookies; wait_for, actions, extract by CSS selector), map, geo compare and a live tab — in a pooled browser, or in a browser and device of your choice on paid plans. Billed like a session (one thread while the call runs, one per country for geo compare); the Community proxy and your own proxies cost nothing extra. Free tools at /functions. | Browser Tools: POST /v1/scrape returns HTML, cleaned HTML, Markdown or readability text, with metadata (title, Open Graph, JSON-LD, status code) and links in every response, and optionally a screenshot and a PDF from the same call. Its options are useProxy and delay — for actions or a specific country the docs point to a session. $5 per 1,000 calls on Launch and Scale, plus proxy traffic; 20 calls a minute per organization. |
| Crawling and bulk | Jobs on paid plans: batch (any number of URLs), crawl and map, with no size or time limit — runs continue past the 24-hour session limit by themselves; results as JSON lines or a zip archive, kept 14 days; stop and resume; signed webhooks. Plus watch (live XHR/WebSocket/DOM events), warmup and monitors (scheduled change detection with diffs). | No crawl, map, batch or scheduling endpoint, and no webhooks documented. The cookbook schedules sessions or scrapes with Convex cron, Temporal or Trigger.dev, and calls a scrape's links "a ready-made frontier for a crawler". |
| Screenshots and PDFs | Screenshot (PNG/JPEG, viewport or full page) and PDF Functions — a text PDF from a Chromium-based browser. | Browser Tools: POST /v1/screenshot (PNG, viewport or full page) and POST /v1/pdf, each returning a hosted file that works without an API key for 7 days; $5 per 1,000 calls. The pricing page names only /scrape on that line; the pricing-and-limits page names all three. |
| Agent tooling | A hosted MCP server at cdpfleet.com/mcp — add it in Claude, ChatGPT or Gemini and sign in with OAuth — or npx cdpfleet-mcp; SDKs for Node.js, Python and Go. | A hosted MCP server (mcp.steel.dev/mcp, OAuth or API key) with scrape and session tools; a Commands API that drives a session over plain HTTP; mouse-and-keyboard computer-use input; a CLI; a large cookbook — Claude, OpenAI and Gemini computer use, browser-use, Stagehand, LangChain, CrewAI, Temporal and more; Files API. |
| Self-hosting | No — a managed fleet. | Yes: steel-browser (Apache 2.0) runs in Docker locally or on your cloud. |
| Where the browser runs | Servers in the EU (Germany and Finland); your proxy decides where traffic exits. | One region, us-east, by the session docs (a scrape recipe still shows other region values); proxies decide where traffic exits. |
| Examples | 89 cases run on production in five languages, all on GitHub with their output. | steel-cookbook: over 90 examples across TypeScript, Python, Go and Rust, mostly agent-framework integrations. |
Steel facts from its public pricing page and documentation, checked on 9 November 2026; list prices in USD. Tell us if something is out of date.
Which one, when
Pick Steel if…
- You build agents and want proxies, captcha solving, Credentials and live view from one API with SDKs — or you want to self-host the same stack.
- You use Selenium.
- You want a page's HTML, Markdown, metadata, screenshot and PDF from one call at a flat $5 per 1,000.
- Your sessions are short (minutes) and you prefer paying $0 up front plus usage.
Pick cdpfleet if…
- You want one-call tasks with actions, a country and a choice of browser, or crawls, batch jobs and monitors without a scheduler of your own.
- Your sessions run longer than Steel's cap for your plan — 15 minutes on Launch, an hour on Scale — or around the clock on dedicated threads.
- Your site must see a specific real browser or version: Chrome beta/dev or a pinned major, Edge, Brave, Firefox, Camoufox, a stealth build on every plan.
- You already pay for proxies and want per-host routing, live swaps and traffic accounting instead of per-GB fees.
- You want per-second metering and Playwright clients in Java, C# or Go as well as Node.js and Python.
- You want to start free — the Community plan includes proxies — and try a browser in live view before writing any code.
Try it
The quickstart is one HTTPS call and five lines of Playwright; the code builder writes the launch for any browser and option set in your language.
Sources
- Steel pricing
- Steel docs: pricing and limits
- Steel docs: session lifecycle
- Steel docs: intro
- Steel docs: session configuration (timeout, proxies, region)
- Steel docs: proxies
- Steel docs: CAPTCHA solving
- Steel docs: live and recorded sessions
- Steel docs: Profiles API
- Steel docs: Credentials API
- Steel docs: Browser Tools (scrape, screenshot, PDF)
- Steel cookbook: scrape
- Steel cookbook: a scheduled price watch (Convex cron)
- Steel docs: MCP server
- Steel docs: Commands API
- steel-dev/steel-browser on GitHub
- steel-dev/steel-cookbook on GitHub