# Browsa: full product and API summary > Browsa (https://browsa.io) is a cloud browser API for AI agents. One API call starts a real browser in the cloud that you drive with your own Playwright, Puppeteer or CDP code, or you describe a task in plain language and an agent runs it in the browser and returns the result as data. Billing is in prepaid credits that never expire; new accounts get 100 free credits after email verification, with no card. Last updated: 2026-11-15. This file summarizes the public pages of browsa.io for language models and AI search. The source of truth is the linked page in each section. Short index: https://browsa.io/llms.txt ## Facts at a glance | Fact | Value | |---|---| | What it is | Cloud browsers for AI agents and automation, with an API, SDKs and an MCP server | | Company | Browsa Technologies LLC | | Website | https://browsa.io | | Dashboard | https://app.browsa.io | | API base URL | https://browsa.io/v1 | | Authentication | `X-API-Key: agt_live_...` or `Authorization: Bearer agt_live_...` | | SDKs | Python `pip install browsa` (Python 3.9+), TypeScript/Node `npm i browsa` (Node 18+), both MIT | | MCP server | `npx -y @browsa/mcp` (Node 18+, MIT) for Claude Desktop, Claude Code, Cursor, VS Code, Cline, Continue, Zed | | OpenAPI spec | https://browsa.io/openapi.yaml (OpenAPI 3.1) | | Browser engine | Phantom, Browsa's own browser, the same on every plan | | Price of a browser | 1 credit per started 12 minutes it runs; 5 credits for a full hour | | Price of a credit | $0.10 (Starter, $20 for 200), $0.083 (Growth, $100 for 1,200), $0.067 (Pro, $500 for 7,500) | | Free credits | 100, added once after email verification. No card. Google sign-ups are verified at once | | Credit expiry | Never | | Subscription | None. Packs stack | | Longest browser session | 1 hour | | Browsers at once | 6 browser slots per workspace (each running browser and each saved identity takes one); more on request | | Proxies | Built-in: datacenter 5 credits per GB, residential 50 credits per GB. Your own HTTP, HTTPS or SOCKS5 proxy per browser is free; without one, traffic uses Browsa's data-center IP address | | CAPTCHA solving | reCAPTCHA v2 image grids (3 x 3, 4 x 4 and reloading tiles), image-text captchas, GeeTest slide puzzles (v3 and v4) and AWS WAF captchas, in hosted agent tasks, on workspaces that bought a pack (fair use 300 an hour, 3,000 a day), solved in the browser's own session; other kinds not yet; types and measured pass rates at https://browsa.io/captcha | | Self-hosting | Not available; Browsa is a hosted service | | SOC 2 / HIPAA | Not today | | Support | support@browsa.io | ## What can you do with Browsa? 1. **Cloud browsers over CDP.** `POST /v1/agents` starts a Phantom browser and returns `cdp_http` (for Playwright and Puppeteer), `cdp_url` (a page WebSocket for raw CDP clients) and `live_url` (a live view you can watch and control). Replace `chromium.launch()` with `chromium.connect_over_cdp(cdp_http, headers={"X-API-Key": API_KEY})`. Guide: https://browsa.io/guides/playwright-puppeteer-cdp 2. **Agent tasks.** `POST /v1/tasks` takes a plain-language `task` and returns a job with the result. Use your own Anthropic, OpenAI or Google key, a key saved in the dashboard, or a hosted tier (`"llm": "browsa"`, `"browsa-fast"` or `"browsa-smart"`) to run with no key. Ask for JSON with `output_schema`. Docs: https://browsa.io/docs#tasks 3. **Quick scrape, extract and screenshot.** `POST /v1/quick/scrape` returns a rendered page as Markdown, HTML, text, links or a screenshot in one call; `/v1/quick/extract` returns fields you describe; `/v1/quick/screenshot` returns a full-page PNG; `/v1/quick/batch` fans out many URLs; `/v1/quick/crawl` follows links from a start page. 1 credit per page. Docs: https://browsa.io/docs#quick 4. **Schedules and webhooks.** Run a task on any interval from 1 hour to 30 days, billed like a normal task, and receive each result on an HMAC-SHA256-signed webhook. Docs: https://browsa.io/docs#schedules 5. **Persistent identities (saved logins).** A persistent browser identity keeps cookies, local storage and fingerprint between sessions. Docs: https://browsa.io/docs#agents-sessions 6. **MCP server.** Gives an AI assistant a cloud browser through MCP. Guide: https://browsa.io/guides/mcp-server ## Example: a real agent run On 14 November 2026 one agent task ran on production in one browser on a hosted tier. The task: open the Show HN page on Hacker News, find the highest-ranked post that links to a GitHub repository, open it, and return JSON. The agent opened news.ycombinator.com/show, clicked rank 3 (the first post linking to GitHub), read github.com/0xdeadf1sh/T1DMAI, went back to confirm the exact title, and returned: ```json {"hn_title": "Show HN: I have trained 15 blood glucose forecasting models on my own data", "hn_points": 36, "repo": "0xdeadf1sh/T1DMAI", "stars": 34, "description": "A model zoo of blood glucose forecasters trained & tested on my own data"} ``` 1 minute 15 seconds from request to result, 8 credits (7 for the task and 5 for the browser, with 4 refunded for unused browser time). A recording of the run, played at 3x speed, is on the home page: https://browsa.io/ ## What is a cloud browser for AI agents? A cloud browser is a real web browser that runs on a provider's servers and is started, driven and stopped through an API. An AI agent or a script uses it to open pages, click, type and read the result without running a browser on its own machine. Explainer: https://browsa.io/guides/what-is-a-cloud-browser ## Quickstart Scrape a page (no agent, no LLM, 1 credit): ``` curl -X POST https://browsa.io/v1/quick/scrape \ -H "X-API-Key: agt_live_..." \ -H "Content-Type: application/json" \ -d '{"url":"https://news.ycombinator.com","formats":["markdown","links"]}' ``` Playwright over CDP (Python): ```python from browsa import Client from playwright.sync_api import sync_playwright API_KEY = "agt_live_..." client = Client(api_key=API_KEY) agent = client.create_agent(ttl="10m") try: with sync_playwright() as p: browser = p.chromium.connect_over_cdp(agent["cdp_http"], headers={"X-API-Key": API_KEY}) context = browser.contexts[0] page = context.pages[0] if context.pages else context.new_page() page.goto("https://example.com") print(page.title()) finally: client.destroy_agent(agent["id"]) ``` Playwright over CDP (TypeScript): ```ts import { Client } from "browsa"; import { chromium } from "playwright"; const apiKey = "agt_live_..."; const client = new Client({ apiKey }); const agent = await client.createAgent({ ttl: "10m" }); const browser = await chromium.connectOverCDP(agent.cdp_http, { headers: { "X-API-Key": apiKey } }); const context = browser.contexts()[0]; const page = context.pages()[0] ?? (await context.newPage()); await page.goto("https://example.com"); ``` Puppeteer: `puppeteer.connect({ browserURL: agent.cdp_http, headers: { "X-API-Key": apiKey } })` with `puppeteer-core`. Do not pass `cdp_url` as `browserWSEndpoint`: it is one page's socket, and Puppeteer fails on it with "Target.getBrowserContexts: Not allowed". Agent task on a hosted tier (no model key): ``` curl https://browsa.io/v1/tasks \ -H "X-API-Key: agt_live_..." \ -H "Content-Type: application/json" \ -d '{"task": "Return the top 3 Hacker News stories as JSON", "llm": "browsa"}' ``` Python SDK: `Client(api_key=...).run_task(task=..., llm="browsa")` returns the finished job (`job.final_result`, `job.live_url`). ## API surface (main endpoints) All paths are under https://browsa.io. Full reference: https://browsa.io/docs and https://browsa.io/openapi.yaml | Endpoint | What it does | |---|---| | `POST /v1/agents` | Start a browser. Burner (TTL 5 minutes to 1 hour) or persistent identity (`"mode": "persistent"`) | | `GET /v1/agents`, `GET /v1/agents/{id}` | List browsers, get one (status, CDP URLs, how it ended) | | `DELETE /v1/agents/{id}` | Stop a browser; unused reserved time is refunded | | `GET /v1/agents/{id}/cdp` | CDP addresses again, to reconnect (409 if not running) | | `POST /v1/agents/{id}/observe`, `POST /v1/agents/{id}/act` | Read the page as compact text with element ids, and perform one action; no CDP client needed, free | | `GET /v1/agents/{id}/downloads`, `GET /v1/agents/{id}/downloads/{name}` | List the files the browser downloaded and fetch one; kept until the browser stops, free | | `POST /v1/agents/{id}/sessions` | Open a session on a persistent identity (1 credit, up to 1 hour) | | `GET /v1/agents/{id}/trust`, `/fingerprint`, `/detection-signal` | Trust score breakdown, fingerprint preview, latest detection-harness record | | `POST /v1/tasks` | Run an agent task from plain language; async, returns a job | | `POST /v1/sessions/{id}/messages`, `POST /v1/sessions/{id}/stop` | Follow-up messages to a task's browser (`keep_browser: true`), and stop it | | `POST /v1/quick/scrape`, `/extract`, `/screenshot`, `/batch`, `/crawl` | One-call page reads | | `POST /v1/auto` | Smart router: a plain fetch skips the model, interactive work goes to an agent | | `GET /v1/jobs/{id}`, `GET /v1/jobs/{id}/events` | Job result and outcome; server-sent events stream | | `POST /v1/schedules` | Recurring tasks | | `/v1/proxies`, `POST /v1/proxies/{id}/check`, `POST /v1/proxies/check` | Saved proxies (you bring the proxy) and an exit report for a saved or unsaved proxy | | `/v1/llm-keys` | Saved Anthropic, OpenAI or Google keys, encrypted, never returned | | `GET /v1/credits`, `GET /v1/credits/ledger` | Balance, packs and ledger | | `GET /v1/budget`, `PUT /v1/budget` | Optional monthly spending limit and alerts | Request details worth knowing: - `POST /v1/agents` body: `ttl` (`"5m"`, `"1h"`, capped at 1 hour), `mode`, `initial_url`, `country` (ISO-2; sets timezone and language when a proxy is set; without a proxy it has no effect, the browser leaves from the data center in Germany and the response says so in `warnings` and `egress`), `os_type` (`OS_TYPE_MACOS` default, `OS_TYPE_WINDOWS`, `OS_TYPE_LINUX`), `proxy` or `proxy_id`, `screen_resolution`, and overrides `webrtc_mode`, `geolocation_mode`, `geolocation_data`, `media_devices_mode`, `port_scan_protection`. - `POST /v1/tasks` body: `task` (required), `llm`, `llm_api_key`, `llm_key_id`, `llm_base_url` (Anthropic- or OpenAI-compatible endpoint), `max_steps` (default 20, max 100), `max_credits`, `output_schema` (JSON Schema, result in `result.structured_output`), `allowed_domains`, `initial_url`, `os_type`, `country`, `secrets`, `credential_ids`, `proxy` or `proxy_id`, `keep_browser`, `browser_ttl_seconds`, `allow_purchase` (default false: the agent refuses order, payment and subscription buttons and the task ends not done, `result.purchase_blocked` says where it stopped). - `Idempotency-Key` header is accepted on `POST /v1/quick/*`, `POST /v1/tasks`, `POST /v1/auto` and session messages (cached 24 hours). `POST /v1/agents` does not accept it. - Every response carries `X-Request-Id`. Errors use one envelope: `{"error": "...", "code": "INSUFFICIENT_CREDIT", "request_id": "..."}`. ## Observe and act without a CDP client For an agent that only has an API key. Both calls work on a running browser and are free. `POST /v1/agents/{id}/observe` returns JSON whose `text` is the page as a person could see and act on it, one element per line: ``` page "Sign up" https://example.com/signup scroll 0 of 713, viewport 1280x713 [1] heading "Sign up" h1 [2] form [3] textbox "Name" empty [4] checkbox "Send me news" [5] button "Greet" [6] link "Second page" → /second dialog: none frames: none ``` Body (all optional): `scope` (`viewport` default, or `page`), `budget` (approximate tokens, default 2000), `expand` (id of a table, list or container to list in full), `tab`, `json` (`true` adds the structure under `view`). Other response fields: `state` (`ok`, `dialog`, or `challenge` when a full-page bot check is up; nothing is read from it), `url`, `title`, `tokens`, `tabs`, `session`, `ids_reset`. `POST /v1/agents/{id}/act` performs one action with real mouse and key events and returns what changed: ``` curl -X POST https://browsa.io/v1/agents/$AGENT_ID/act \ -H "X-API-Key: agt_live_..." -H "Content-Type: application/json" \ -d '{"action": "type", "id": 3, "text": "Ada"}' after: type [3] "Ada" ~ [3] textbox "Name" value "Ada" settled after 0 ms, no requests ``` `~` changed, `+` added, `-` removed. After a navigation or a scroll, `text` carries a new view instead. | Action | Fields | |---|---| | `click` | `id`, or `role` + `name`, or `x` + `y`; optional `double`, `right`, `modifiers` | | `type` | `id`, `text`; optional `clear`, `submit` (presses Enter) | | `press` | `keys`: `Enter`, `Control+A`, `Tab Tab Enter` | | `select` | `id`, `option` (text or value) | | `check`, `uncheck`, `hover` | `id` | | `scroll` | `dy` (pixels), or `to` (`top`, `bottom`), or `id` | | `navigate` | `url` | | `back`, `forward`, `reload` | none | | `wait` | `ms`, or `for`: `text` with `text`, `id-gone` with `id`, `url-change`; optional `timeout`; at most 30 seconds | | `upload` | `id`, `files` (paths on the browser's machine) | | `dialog` | `accept`, optional `text`; answers a JavaScript dialog | - The HTTP status is 200 whenever the browser was reached. Read `status`: `ok`, `failed` (with `reason`), `blocked`, `challenge` or `stale`. - An id is valid from the view that listed it until the page loads a new document. Browsa keeps the view for 10 minutes after the last call. When it had to start a new one, the response has `ids_reset: true`, and an action that names an id is not performed: `status` is `stale` and `text` carries a new view. Send the `session` value of the last response with each call to have this checked. - One call at a time per browser: a second call waits up to 30 seconds, then gets `409 BROWSER_BUSY`. A CDP client can stay connected at the same time. ## How do cloud browsers on Browsa differ from a local browser? - Screen size, user agent, timezone, language, geolocation and CPU count are set together when the browser starts. CDP emulation commands that change one of them alone (for example Playwright `page.set_viewport_size()`, Puppeteer `page.setViewport()` and `page.setUserAgent()`) return success and change nothing. Set `screen_resolution` and `os_type` at creation, or add `cdp_overrides=allow` to the CDP address to forward emulation commands. - Files a page downloads stay in the cloud browser; fetch them from inside the page and return the bytes. - The browser starts with a context and an open page; use `browser.contexts[0]`. - A browser ends at its TTL (at most 1 hour) or when you destroy it, not on idleness. Closing the CDP connection does not stop it. - Playwright and Puppeteer enable CDP's Runtime domain when they attach, which some bot checks can see. ## Pricing Source: https://browsa.io/pricing and https://browsa.io/pricing.md | What you run | Credits | |---|---| | Cloud browser | 1 per started 12 minutes it runs, 5 for a full hour | | Quick scrape, extract or screenshot | 1 per page | | Batch of URLs | 1 per URL | | Agent task, your own LLM key | browser + 1 + 1 per 5 steps | | Agent task on a hosted tier (`"llm": "browsa"`, `"browsa-fast"` or `"browsa-smart"`) | browser + 1 + 1 per 5 steps + 5 | | Persistent identity | 50 per 30 days | | Session on a persistent identity | 1 (up to 1 hour) | | Webhooks, job status, polling, downloads | Free | Credit packs: Starter $20 for 200 credits, Growth $100 for 1,200 credits (20% bonus), Pro $500 for 7,500 credits (50% bonus). Paid by card, Apple Pay or Google Pay through Stripe; invoicing on request. Minimum purchase is one Starter pack. Worked examples: a 10-step agent task that finishes inside 12 minutes costs 4 credits with your own key, 9 on a hosted tier. A 20-step task costs 6 or 11. 1,000 short scrapes cost 1,000 credits ($100 at the Starter rate, $66.67 at the Pro rate). One persistent identity used once a day for a month costs 80 credits. When a browser starts, Browsa holds the price of its TTL (1 credit up to 5 minutes, 5 credits above that) and refunds the unused 12-minute blocks when it ends. A browser that fails to start is refunded. A task that fails gets its whole browser charge back and is not charged for steps or the hosted surcharge. With your own model key, Browsa adds no LLM markup; your provider bills you for tokens. ## Limits | Limit | Value | |---|---| | Browser slots per workspace | 6; a request with every slot in use returns `429 TOO_MANY_REQUESTS` with `Retry-After: 30` and is not charged | | Burner TTL | 5 minutes to 1 hour | | Session on a persistent identity | up to 1 hour | | Agent task steps | default 20, max 100 | | Balance at 0 | `402 INSUFFICIENT_CREDIT` | | Monthly spending limit | optional, 1 to 1,000,000 credits; `402 BUDGET_REACHED` when reached; resets on the 1st (UTC) | | Saved proxies | up to 50 per workspace | ## Proxies Bring your own proxy, free, or use Browsa's built-in proxy network, paid per GB (datacenter 5 credits per GB, residential 50). Without a proxy, every browser uses the IP address of Browsa's data center, with the data center's timezone and English. Pass your own HTTP, HTTPS or SOCKS5 proxy in `proxy` (or a saved one in `proxy_id`) on `POST /v1/agents`, `POST /v1/tasks`, the quick endpoints and schedules; the browser's timezone and language then follow the proxy's country. Proxy traffic costs nothing extra on Browsa. Check a proxy before you spend credits: `POST /v1/proxies/{id}/check` (saved) or `POST /v1/proxies/check` with `{"proxy": {...}}` (not saved). Free, up to 20 seconds, 20 per workspace per hour. Through the proxy it reports `reachable`, `verdict` (`good`, `usable`, `poor`), the types the port answers as (`protocols.answers`), the exit address, where address databases place it and whether they agree (`location.agree`), the time zone the browser will take (`location.chosen_timezone`), whether the address is listed as hosting, proxy, VPN, Tor or mobile (`classification`), and a short load test (`load`: 30 parallel tunnels, then 6 seconds at 5 requests a second, refusals counted by kind). `warnings` codes: `LOCATION_SOURCES_DISAGREE`, `PROXY_REFUSES_UNDER_LOAD`, `FLAGGED_AS_HOSTING`, `FLAGGED_AS_PROXY`, `DECLARED_TYPE_NOT_ANSWERED`, `COUNTRY_MISMATCH`, `EXIT_IP_CHANGES`, `SLOW_PROXY`, `CHECK_INCOMPLETE`. Your proxy password goes only to your proxy; the lookup services (ipinfo, Cloudflare, GeoJS, ipquery, proxycheck) are asked through the proxy and see its exit address. ## Which model runs an agent task? Set `llm` to a hosted tier and send no model key, or bring your own key from Anthropic, OpenAI, Google or any OpenAI-compatible server (`llm_base_url`). With your own key there is no hosted charge; you pay your provider directly. Source: https://browsa.io/guides/models | Tier (`llm`) | Runs | Falls back to | Reasoning | Credits per task | |---|---|---|---|---| | `browsa` (default) | DeepSeek V4.1 Flash | MiniMax M3.1 Flash, then Claude Sonnet 4.6 | No fixed level | 5 | | `browsa-fast` | DeepSeek V4.1 Flash | MiniMax M3.1 Flash, then Gemini 3.1 Flash Lite | No fixed level | 5 | | `browsa-smart` | Kimi K3.1 Pro | Claude Opus 4.8, then Claude Sonnet 4.6 | No fixed level | 5 | - The tier credits are added once per task, on top of the normal task credits (browser time, 1 base credit, 1 credit per 5 steps). They are a server setting and can change; `GET /v1/models` returns the live values. - `reasoning` (optional): `"low"`, `"medium"` or `"high"`. On a hosted tier the credits per task do not change with the level. - A finished job reports `llm_tier`, the reasoning level applied, token usage and, for your own key, an estimated model cost. - A model that cannot run the task fails before a browser starts, with nothing charged. A missing key returns `400 LLM_KEY_REQUIRED`. ## Model keys and security - Your Anthropic, OpenAI or Google key is encrypted before it is stored, opened only by the worker running your task, and deleted when the task ends. It is never logged or shown back. - Saved LLM keys are encrypted and never returned by any endpoint (responses show a masked key). Saved proxy passwords are encrypted with AES-256-GCM and never returned. - API keys are stored as a SHA-256 hash; the raw secret is shown once. - If you run the agent yourself and connect over CDP, your model key never reaches Browsa. ## MCP server `@browsa/mcp` runs with `npx -y @browsa/mcp` and needs `BROWSA_API_KEY`. Tools: `run_task`, `get_job`, `cancel_job`, `respond_to_job`, `list_job_files`, `create_schedule`, `list_schedules`, `list_llm_keys`, `credits`, `list_webhooks`, `create_webhook`, `delete_webhook`, `list_identities`. With no saved model key and no model named, tasks fall back to the default hosted tier, `browsa` (5 extra credits), and the result says so. Claude Code: `claude mcp add browsa -e BROWSA_API_KEY=agt_live_... -- npx -y @browsa/mcp` Claude Desktop and Cursor: ```json { "mcpServers": { "browsa": { "command": "npx", "args": ["-y", "@browsa/mcp"], "env": { "BROWSA_API_KEY": "agt_live_..." } } } } ``` ## Phantom and measured results Browsa runs Phantom, its own browser, for every customer on every plan. It reports `Chrome` (not `HeadlessChrome`) in the user agent, `navigator.webdriver` as `false`, and a GPU name in Chrome's format for the chosen OS (drawing still runs in software, a known gap). The changes are compiled into the browser; no scripts are injected into pages. Details and gaps: https://browsa.io/phantom Dated results for a default browser with no proxy, measured on 14 November 2026 (one run; source: https://browsa.io/benchmarks): - bot.sannysoft.com: 31 of 31 checks passed. - BotD: not detected. browserscan.net: passed. iphey.com: "Trustworthy", score 100. - CreepJS: mixed (0% headless, 0% stealth, two WebGL notes). pixelscan.net: no automated behavior, proxy detected for the data-center address, fingerprint "masking detected". - rebrowser-bot-detector: all 18 tests it ran passed, with unmodified Playwright as the client. - deviceandbrowserinfo.com: not flagged (0 of 22 signals). - fingerprint-scan.com: bot score 0 of 100 (lower is better). - Twelve real sites: 7 loaded with no interaction, 3 showed Cloudflare's full-page check and opened after one click on its checkbox, 2 did not load. No browser passes every check. Results on your own proxy, OS or saved identity can differ. The `trust_score` returned with each browser is an estimate computed from its settings, not a detection result. ## How does Browsa compare? Summaries of the sourced comparison pages. Competitor figures were read on the dates shown on each page. - Browsa vs Browserbase (https://browsa.io/vs/browserbase): Browserbase has monthly plans, cheaper browser-hours at steady volume ($0.10 to $0.12 overage), sessions up to 6 hours on paid plans, included proxies and CAPTCHA solving, SOC 2 Type II and a HIPAA BAA on Scale. Browsa has no monthly fee, credits that never expire, its Phantom browser on every pack, and self-serve hosted agent tasks on your own model key. - Browsa vs Steel (https://browsa.io/vs/steel): Steel is open source (Apache-2.0, self-hostable) and has the cheapest browser-hour ($0.08 to $0.10, by the minute). Browsa costs more per hour and includes one-hour sessions, its Phantom browser, schedules and hosted agent tasks without a monthly plan. - Browsa vs Browser Use Cloud (https://browsa.io/compare): Browser Use Cloud lists residential proxies, automatic CAPTCHA solving, stealth browsers, profiles, webhooks and hosted agents, pay as you go. Browsa supplies no proxies and solves no CAPTCHAs. - All four side by side: https://browsa.io/vs/browserbase-alternatives ## Frequently asked questions **Will my Playwright code work?** Yes. Replace `chromium.launch()` with `chromium.connect_over_cdp(cdp_http, headers={"X-API-Key": API_KEY})` and drop headless and user-agent flags. Puppeteer and raw CDP work the same way. **Do I need my own AI model key?** No. Bring an Anthropic, OpenAI or Google key with no markup, or set `"llm": "browsa"` for a hosted tier, 5 extra credits a task. See the Models guide. (https://browsa.io/guides/models) **Does Browsa provide proxies?** Yes. Datacenter at 5 credits per GB, residential at 50 credits per GB. Or bring any HTTP, HTTPS or SOCKS5 proxy per browser, free. Without one, browsers connect from Browsa's data center. **Does Browsa solve CAPTCHAs?** Partly. Hosted agent tasks solve reCAPTCHA v2 image grids (3 x 3, 4 x 4 and reloading tiles), image-text captchas, GeeTest slide puzzles (v3 and v4) and AWS WAF captchas in the browser's own session, on workspaces that have bought a pack, at no extra charge, with a fair-use limit of 300 solves an hour and 3,000 a day. Other kinds (Enterprise, hCaptcha, Arkose) are not solved yet. The same solver answers the page-view action `solve_captcha` and, in your own Playwright or Puppeteer code, the CDP command `Browsa.solveCaptcha` (events `Browsa.captchaSolving`, `Browsa.captchaSolved`, `Browsa.captchaFailed`). Phantom passes many checks without a captcha being shown at all. Which types are solved and how often the site's own server accepted the answer, with recordings and logs: https://browsa.io/captcha **Do credits expire?** No. **What happens if a browser fails to start or a task fails?** A browser that fails to start is refunded automatically. A failed task gets its whole browser charge back and is not charged for steps. **How long can a browser run?** Up to 1 hour. Persistent identities keep logins between sessions. **How many browsers can I run at once?** 6 browser slots per workspace; email support@browsa.io for more. **Can I self-host Browsa?** No. It is a hosted service only. **Does Browsa have SOC 2 or HIPAA?** Not today. **Can I watch a run?** Yes. Browsers and tasks return a `live_url` that opens the running browser; you can watch and use the mouse and keyboard. **What if a site blocks my agent?** You see the block page in your results and your code decides what to do: retry with a new browser, switch proxy or stop. ## Guides and pages - Guides: https://browsa.io/guides - What is a cloud browser for AI agents: https://browsa.io/guides/what-is-a-cloud-browser - Playwright and Puppeteer over CDP: https://browsa.io/guides/playwright-puppeteer-cdp - MCP server setup: https://browsa.io/guides/mcp-server - browser-use on Browsa: https://browsa.io/guides/browser-use - Stagehand on Browsa: https://browsa.io/guides/stagehand - OpenAI Agents SDK on Browsa: https://browsa.io/guides/openai-agents-sdk - Claude computer use on Browsa: https://browsa.io/guides/claude-computer-use - Use cases: https://browsa.io/use-cases - Docs: https://browsa.io/docs - Cookbook: https://browsa.io/cookbook - Pricing: https://browsa.io/pricing (Markdown: https://browsa.io/pricing.md) - Phantom: https://browsa.io/phantom - Benchmarks: https://browsa.io/benchmarks - Status: https://browsa.io/status - Terms: https://browsa.io/terms - Privacy: https://browsa.io/privacy - Contact: support@browsa.io