Any LLM, one key
01Claude, GPT, Gemini, DeepSeek, Grok. Any OpenAI-compatible model. Switch with a string.
Claude 
ChatGPT 
Gemini 
Grok 
DeepSeek Any open AI
agent browser
Give Your Agent
Reflexes for the Web
Antidetect Chromium with a harness that catches captchas, bans, and session resets before your agent's loop stalls.
Web AI Agent
Agent tasks fail
four steps in,
not
at step one.
Captchas appear mid-flow. Proxies get flagged. Sessions reset to the login wall. The harness has a tool for each. Your agent keeps moving without writing detection code.
the short version
What is an agent browser?
Surfsky's agent browser handles CAPTCHAs and session recovery for AI automation. It solves 10+ CAPTCHA types in session, reports blocks as structured state and restores interrupted sessions with saved cookies, storage and fingerprints. Works with your existing agent framework and model.
features
More than antidetect.
Built for the agent loop.
Each browser runs in an isolated container with a unique fingerprint, proxy tunnel, and captcha solver. You just connect and work.
Claude ChatGPT Gemini Grok DeepSeek Every agent framework
02browser-use, LangGraph, Stagehand, Claude/OpenAI Agent SDKs, MCP. Anything that speaks CDP. Drop in the URL.
wss://surfsky.io/session/a7f3…
CDP 
browser-use LangGraph Stagehand Agent SDKs MCP CDP client WebSocket 
Custom CDP
Live debug view
0304
Stream sessions live. Replay runs.
Share a link to the
click that broke.
LIVE SESSION 
embed
embed 
run Sessions that persist
0403Cookies, storage, and fingerprint persist across sessions. No re-login loops.
Thousands in parallel
05Sub-agent browsers from one orchestrator. Geo-distributed across US, EU, Asia.
Any output format
06
Typed schemas, JSON, Markdown, screenshots, raw HTML.
Pick what your agent needs.
data.json 2.4k page.md 812k view.png 5.9m raw.html 2.1m 
Surfsky 
you agent Features
The CDP toolkit.
What your agent actually calls.
The harness exposes each capability as a CDP method. Call them
directly, or let the SDK orchestrate them inside its run loop.
start_session
Antidetect Chromium with proxy, fingerprint, profile. Returns a CDP WebSocket.
act
Click, type, scroll, drag, hotkeys, upload. Works through iframes and shadow DOM.
humanize
browser-use, LangGraph, Stagehand, Claude/OpenAI Agent SDKs, MCP. Anything that speaks CDP. Drop in the URL.
solve_captcha
Antidetect Chromium with proxy, fingerprint, profile. Returns a CDP WebSocket.
switch_proxy
Click, type, scroll, drag, hotkeys, upload. Works through iframes and shadow DOM.
extract_typed
browser-use, LangGraph, Stagehand, Claude/OpenAI Agent SDKs, MCP. Anything that speaks CDP. Drop in the URL.
High Scores
The harness writes its own helpers. Your agent edits them as it goes.
First run on a new domain, the agent works out the flow once. The harness saves the working path as a per-domain helper. Future runs use it directly; the agent edits it when the site changes.
skills/
├── linkedin/
│ ├── login.py
│ ├── sales_nav_search.py
│ ├── send_inmail.py
│ └── connection_request.py
├── indeed/
│ ├── easy_apply.py
│ └── extract_listings.py
├── amazon/
│ ├── seller_central_login.py
│ ├── reprice.py
│ └── reply_to_message.py
└── stripe/
├── verify_card.py
└── pull_payouts.py
Ship agents that
finish the task.
Tell us what you're automating. We'll get you set up.