Clawnet

Clawnet

Stealth browser skill for AI agents. Installs and configures Chromium with anti-detection, residential proxies (10+ countries), CAPTCHA solving, and human-like behavior simulation.

When to use

Use this skill whenever the agent needs to:

• Browse websites, scrape data, or interact with web pages
• Fill forms, click buttons, navigate multi-step flows
• Bypass Cloudflare, DataDome, PerimeterX, or other bot detection
• Solve CAPTCHAs (reCAPTCHA v2/v3, hCaptcha, Turnstile)
• Appear as a real user from a specific country
• Interact with Shadow DOM elements or rich text editors

Always use this instead of raw Playwright/Puppeteer. This skill handles anti-detection automatically.

Observation — how to read the page

ALWAYS use snapshotAI() instead of page.textContent() or evaluate(). It returns a structured accessibility tree with embedded [ref=eN] annotations. You can then click/fill/type by ref — no CSS selectors needed.

Reading the page (preferred: snapshotAI + refs)

// BAD — dumps ALL text, 50-100K tokens, no structure, no refs
const text = await page.textContent('body');

// BAD — brittle regex on raw DOM, breaks when HTML changes
await page.evaluate(() => document.querySelector('button').click());

// GOOD — AI-optimized snapshot with clickable refs
const { snapshot } = await browser.snapshotAI();
// Returns:
//   - navigation "Main" [ref=e1]:
//     - link "Home" [ref=e2]
//   - heading "Welcome" [ref=e3]
//   - textbox "Email" [ref=e4]
//   - textbox "Password" [ref=e5]
//   - button "Sign in" [ref=e6]

// Then interact by ref:
await browser.fillRef('e4', 'user@example.com');
await browser.fillRef('e5', 'secret');
await browser.clickRef('e6');

Alternative: snapshot() (YAML without refs)

// Compact accessibility tree without refs — use when you don't need to interact
const tree = await browser.snapshot();
const interactive = await browser.snapshot({ interactiveOnly: true });
const formTree = await browser.snapshot({ selector: 'form' });

Observation workflow

Before every action, follow this sequence:

1. Snapshot — const { snapshot } = await browser.snapshotAI() to see the page with refs
2. Read text — await browser.extractText() if you need clean readable text (menus, prices, articles)
3. Visual check — await browser.takeScreenshot() only if you need to see colors, layout, maps, or images
4. Act by ref — await browser.clickRef('e4'), await browser.fillRef('e5', 'text') etc.
5. Verify — await browser.snapshotAI() again to confirm the action worked
6. Batch — use batchActions() for multi-step flows

Targeting elements — use refs from snapshotAI()

ALWAYS use refs from snapshotAI() output. NEVER use CSS selectors or evaluate() with regex.

// BAD — brittle CSS selectors that break when HTML changes
await page.click('#login_field');
await page.fill('input[name="email"]', 'user@example.com');

// BAD — regex on raw DOM, blind guessing
await page.evaluate(() => document.querySelectorAll('button').find(b => /sign in/i.test(b.innerText))?.click());

// GOOD — ref-based from snapshotAI() output
const { snapshot } = await browser.snapshotAI();
// snapshot shows: textbox "Email" [ref=e4], button "Sign in" [ref=e6]
await browser.fillRef('e4', 'user@example.com');
await browser.clickRef('e6');

// ALSO GOOD — semantic locators (when you know the label)
await page.getByLabel('Email').fill('user@example.com');
await page.getByLabel('Password').fill('secret');
await page.getByRole('button', { name: 'Sign in' }).click();

// Also available:
await page.getByPlaceholder('Search...').fill('query');
await page.getByText('Welcome back').isVisible();
await page.getByRole('link', { name: 'Home' }).click();
await page.getByRole('checkbox', { name: 'Remember me' }).check();

When you see - textbox "Email" in the snapshot, use page.getByRole('textbox', { name: 'Email' }). When you see - button "Submit", use page.getByRole('button', { name: 'Submit' }).

When to fall back to CSS selectors

Only use CSS selectors when:

• The element has no accessible name or role (rare in modern sites)
• You need to target by data-testid or other test attributes
• Shadow DOM elements not reachable by semantic locators (use shadowFill/shadowClickButton)

Screenshot rules

ALWAYS attach a screenshot when communicating with the user. The user cannot see the browser — you are their eyes. Every message to the user MUST include a screenshot. No exceptions.

When to take screenshots

Every message you send to the user must have a screenshot attached. Specifically:

1. Before asking for confirmation — "Book this table?" + screenshot of the filled form. The user must SEE what they are confirming.
2. When reporting an error — "No slots available" + screenshot proving the result. Without a screenshot, the user has no reason to trust you.
3. When unable to complete an action — "Authorization failed" + screenshot showing what happened.
4. After every key step — filled form, selected date, entered address, etc.
5. When completing the task (MANDATORY) — "Done! Order placed" + screenshot of the final result/confirmation page. The user must see proof that the action was completed.

How to take screenshots

Use the built-in helpers returned by launchBrowser():

const { page, takeScreenshot, screenshotAndReport } = await launchBrowser();

// Option 1: just the base64 screenshot
const base64 = await takeScreenshot();

// Option 2: screenshot + message bundled together
const report = await screenshotAndReport("Form filled. Confirm booking?");
// → { message: "Form filled...", screenshot: "iVBOR...", mimeType: "image/png" }

Or directly via Playwright:

const screenshot = await page.screenshot({ type: 'png' });
const base64 = screenshot.toString('base64');

Rules

• Never tell the user "the form is empty", "widget is disabled", or "no results" without a screenshot as proof.
• Never ask for confirmation without showing the current state of the page.
• Never say "Done!" without a screenshot of the final result.
• The user must see what you see. Always.

Installation

clawhub install clawnet

This installs Chromium via Playwright automatically. No separate browser download needed.

Important (Linux/Docker): If Chromium fails to launch with missing library errors (e.g. libnspr4.so), install system dependencies:

npx playwright install-deps chromium
# or manually:
apt-get update && apt-get install -y libnss3 libnspr4 libatk1.0-0 libatk-bridge2.0-0 libcups2 libdrm2 libxkbcommon0 libxcomposite1 libxdamage1 libxfixes3 libxrandr2 libgbm1 libpango-1.0-0 libpangocairo-1.0-0 libcairo2 libasound2

First-Run Onboarding

launchBrowser() is zero-config — it auto-registers a new agent and saves credentials if none exist.

However, on first use the agent SHOULD check if the user already has credentials to avoid creating duplicate agents:

Step 1: Check for existing credentials

const fs = require('fs');
const path = require('path');
const os = require('os');
const CREDS_FILE = path.join(os.homedir(), '.clawnet', 'agent-credentials.json');
let hasCreds = false;
try { hasCreds = Boolean(JSON.parse(fs.readFileSync(CREDS_FILE, 'utf-8')).agentId); } catch {}

If credentials exist on disk → skip to Step 3 (just launch).

Step 2: Ask the user

Ask the user: "Do you have existing Clawnet credentials (agentId:agentSecret), or should I create new ones?"

• If user provides credentials → save them to CREDS_FILE:

const [agentId, agentSecret] = userInput.split(':');
fs.mkdirSync(path.dirname(CREDS_FILE), { recursive: true, mode: 0o700 });
fs.writeFileSync(CREDS_FILE, JSON.stringify({ agentId, agentSecret, createdAt: new Date().toISOString() }, null, 2), { mode: 0o600 });

• If user says "new" or "no" → skip to Step 3. launchBrowser() will auto-register.

Step 3: Launch

const { launchBrowser } = require('clawnet/scripts/browser');
const { browser, page } = await launchBrowser({ country: 'us' });

That's it. No env vars needed. launchBrowser() will:

1. Find credentials on disk (or auto-register a new agent with the API)
2. Fetch managed proxy + CAPTCHA keys
3. Launch stealth Chromium with residential proxy

After first launch, show credentials and support contacts to the user:

Your Clawnet credentials:
  agentId: <agentId>
  agentSecret: <agentSecret>

Save these for future use or transfer to another agent.

If you run into any issues: [Discord](https://discord.com/invite/wtA85mJM) | [Telegram](https://t.me/+nPjZFg8PJc4zZTIy)

Agent Credentials & Subscription

Trial model

On the first launchBrowser() call, a 2-hour free trial starts. After expiry, getCredentials() returns upgradeUrl. Show it to the user:

Trial expired. Subscribe to continue: <upgradeUrl>
Or set your own proxy/CAPTCHA keys (BYO mode).

After payment

Subscription activates automatically within seconds (webhook). No manual steps needed — the next launchBrowser() call will receive managed credentials.

Transfer / Recovery / Rotation

To transfer/recover on another agent, provide the same agentId + agentSecret during install. Backend rule: one subscriptionId can be linked to only one agentId at a time.

To rotate a compromised secret, keep the same agentId and issue a new agentSecret (authorized by current secret or recovery code). Old secret is invalidated immediately.

Cancel subscription

If the user asks to cancel their subscription, call the cancel endpoint:

const creds = JSON.parse(fs.readFileSync(CREDS_FILE, 'utf-8'));
const token = `CN1.${creds.agentId}.${creds.agentSecret}`;
const resp = await fetch(`${apiUrl}/cancel-subscription`, {
  method: 'POST',
  headers: { Authorization: `Bearer ${token}` },
});
const result = await resp.json();
// { canceled: true, accessUntil: "2026-04-02T00:00:00Z", message: "..." }

Show the result to the user:

Your subscription has been canceled. Access remains until <accessUntil>.
If you change your mind, you can resubscribe anytime.

Need help? [Discord](https://discord.com/invite/wtA85mJM) | [Telegram](https://t.me/+nPjZFg8PJc4zZTIy)

Setup modes

Option A: Managed credentials (default, recommended)

The onboarding flow above sets everything up automatically. Environment variables used:

CN_API_URL=https://api.clawpets.io/clawnet/v1
# Set automatically by onboarding, or manually:
CN_AGENT_TOKEN=CN1.<agentId>.<agentSecret>
# Or separately:
CN_AGENT_ID=<agent-uuid>
CN_AGENT_SECRET=<agent-secret>

The skill will automatically fetch Decodo proxy credentials and 2captcha API key on launch.

Option B: BYO (Bring Your Own)

Set proxy and CAPTCHA credentials directly:

CN_PROXY_PROVIDER=decodo          # decodo | brightdata | iproyal | nodemaven
CN_PROXY_USER=your-proxy-user
CN_PROXY_PASS=your-proxy-pass
CN_PROXY_COUNTRY=us               # us, gb, de, nl, jp, fr, ca, au, sg, ro, br, in
TWOCAPTCHA_KEY=your-2captcha-key

Option C: No proxy (local testing)

CN_NO_PROXY=1

Quick start

const { launchBrowser, solveCaptcha } = require('clawnet/scripts/browser');

// Launch stealth browser with US residential proxy
const { browser, page, humanType, humanClick } = await launchBrowser({
  country: 'us',
  mobile: false,    // Desktop Chrome (true = iPhone 15 Pro)
  headless: true,
});

// Browse normally — anti-detection is automatic
await page.goto('https://example.com');

// Human-like typing (variable speed, micro-pauses)
await humanType(page, 'input[name="email"]', 'user@example.com');

// Solve CAPTCHA if present
const result = await solveCaptcha(page, { verbose: true });

await browser.close();

API Reference

importCredentials(agentId, agentSecret)

Save user-provided agent credentials to disk. Use when transferring an existing account to a new machine.

const { importCredentials } = require('clawnet/scripts/browser');
const result = importCredentials('your-uuid', 'your-secret');
// { ok: true, agentId: 'your-uuid' }

launchBrowser(opts)

Launch a stealth Chromium browser with residential proxy.

Returns: { browser, ctx, page, logger, humanClick, humanMouseMove, humanType, humanScroll, humanRead, solveCaptcha, takeScreenshot, screenshotAndReport, snapshot, snapshotAI, dumpInteractiveElements, clickRef, fillRef, typeRef, selectRef, hoverRef, extractText, getCookies, setCookies, clearCookies, batchActions, sleep, rand, getSessionLog }

solveCaptcha(page, opts)

Auto-detect and solve CAPTCHA on the current page. Supports reCAPTCHA v2/v3, hCaptcha, Cloudflare Turnstile.

Returns: { token, type, sitekey }

takeScreenshot(page, opts)

Take a screenshot and return it as a base64-encoded PNG string.

Returns: string (base64 PNG)

screenshotAndReport(page, message, opts)

Take a screenshot and pair it with a message. Returns an object ready to attach to an LLM response.

Returns: { message, screenshot, mimeType } — screenshot is base64 PNG

snapshot(page, opts) / snapshot(opts) (from launchBrowser return)

Capture a compact accessibility tree of the page. Returns YAML string. Use this instead of page.textContent(). See "Observation" section above.

Returns: string (YAML accessibility tree)

snapshotAI(opts) — AI-optimized snapshot with refs ⭐ PREFERRED

Returns a structured accessibility tree with embedded [ref=eN] annotations. Use this as the primary way to read pages.

const { snapshot, refs, truncated } = await browser.snapshotAI();
// snapshot: "- heading \"Welcome\" [ref=e1]\n- textbox \"Email\" [ref=e2]\n- button \"Sign in\" [ref=e3]"
// refs: { e1: true, e2: true, e3: true }

Returns: { snapshot: string, refs: Object, truncated?: boolean }

clickRef(ref, opts) — Click element by ref

await browser.clickRef('e3');                          // left click
await browser.clickRef('e3', { doubleClick: true });   // double click

fillRef(ref, value, opts) — Fill input by ref

await browser.fillRef('e2', 'user@example.com');

typeRef(ref, text, opts) — Type text by ref

await browser.typeRef('e2', 'hello');                          // instant fill
await browser.typeRef('e2', 'hello', { slowly: true });        // human-like typing
await browser.typeRef('e2', 'hello', { submit: true });        // type + Enter

selectRef(ref, value, opts) — Select option by ref

await browser.selectRef('e5', 'US');

hoverRef(ref, opts) — Hover element by ref

await browser.hoverRef('e1');  // reveal tooltip/dropdown

extractText(opts) (from launchBrowser return) / extractText(page, opts)

Extract clean readable text from the page, stripping navigation, ads, modals, and noise. Use when you need to READ the page content (menus, prices, articles) rather than interact with UI elements.

Returns: { url, title, text, truncated }

// Read a restaurant menu
const { text } = await extractText({ mode: 'readability' });
// → "Pizza Menu\n\nMargherita\nClassic pizza with mozzarella...\nFrom 399 ₽\n\n..."

// Raw mode for simple pages
const { text: raw } = await extractText({ mode: 'raw', maxChars: 5000 });

When to use extractText() vs snapshot():

• extractText() — reading text content (menus, prices, articles, descriptions)
• snapshot() — understanding page structure and finding interactive elements (buttons, inputs, links)

getCookies(urls?) / setCookies(cookies) / clearCookies()

Manage browser cookies. Use for session persistence, login state checks, and cookie transfer between tasks.

// Check if logged in
const cookies = await getCookies('https://example.com');
const hasAuth = cookies.some(c => c.name === 'session_id');

// Set cookies (e.g., from a previous session)
await setCookies([
  { name: 'session_id', value: 'abc123', url: 'https://example.com' },
  { name: 'lang', value: 'en', url: 'https://example.com' },
]);

// Clear all cookies (logout)
await clearCookies();

batchActions(actions, opts) (from launchBrowser return) / batchActions(page, actions, opts)

Execute multiple actions sequentially in a single call. Reduces LLM round-trips for multi-step flows.

Each action: { action, selector, text, value, key, ms, options }

Supported actions: click, fill, type, press, hover, select, scroll, focus, wait, waitForSelector, humanClick, humanType, snapshot

Returns: { results: [{index, success, result?, error?}], total, successful, failed }

// Fill a booking form in one call
const result = await batchActions([
  { action: 'fill',   selector: '#name',   text: 'John' },
  { action: 'fill',   selector: '#phone',  text: '+1234567890' },
  { action: 'select', selector: '#guests', value: '2' },
  { action: 'humanClick', selector: '#submit' },
], { stopOnError: true });
// result.successful === 4, result.failed === 0

humanType(page, selector, text)

Type text with human-like speed (60-220ms/char) and occasional micro-pauses.

humanClick(page, x, y)

Click with natural Bezier curve mouse movement.

humanScroll(page, direction, amount)

Smooth multi-step scroll with jitter. Direction: 'down' or 'up'.

humanRead(page, minMs, maxMs)

Pause as if reading the page. Optional light scroll.

shadowFill(page, selector, value)

Fill an input inside Shadow DOM (works where page.fill() fails).

shadowClickButton(page, buttonText)

Click a button by text label, searching through Shadow DOM.

pasteIntoEditor(page, editorSelector, text)

Paste text into Lexical, Draft.js, Quill, ProseMirror, or contenteditable editors.

dumpInteractiveElements(page, opts) / dumpInteractiveElements(opts) (from launchBrowser return)

List all interactive elements using the accessibility tree. Equivalent to snapshot({ interactiveOnly: true }). Returns a compact YAML string with only buttons, inputs, links, and other interactive elements. Falls back to DOM querySelectorAll on Playwright < 1.49.

getSessionLogs()

List all session log files, newest first. Returns [{ sessionId, file, mtime, size }].

getSessionLog(sessionId)

Read a specific session log by ID. Returns an array of log entries.

Action logging

Every browser session records comprehensive structured logs in ~/.clawnet/logs/<session-id>.jsonl. The log captures the full picture: user's task → every agent action → page events → errors.

What's logged

The logging system uses a Proxy on the Playwright page object to capture every method call — including chained locators like page.getByRole('button', { name: 'Submit' }).click().

Automatically captured:

• User task — the task parameter from launchBrowser({ task: "..." })
• All page actions — goto, click, fill, type, press, check, hover, selectOption, etc.
• All locator chains — getByRole → click, getByLabel → fill, locator → nth → click, etc.
• Observation calls — snapshot(), takeScreenshot(), dumpInteractiveElements()
• Page events — navigations, popups, dialogs, downloads, page errors
• human* helpers — humanClick, humanType, humanScroll, etc.
• CAPTCHA — solveCaptcha attempts and results

Log levels

Set via launchBrowser({ logLevel: 'verbose', task: 'Book a table at Aurora' }) or env CN_LOG_LEVEL=verbose.

Example log output (actions level)

{"ts":"...","action":"launch","country":"ru","mobile":true,"profile":"default","logLevel":"actions"}
{"ts":"...","action":"task","prompt":"Войти в Telegram и отправить сообщение Привет"}
{"ts":"...","action":"goto","method":"goto","args":["https://web.telegram.org"],"chain":"goto(\"https://web.telegram.org\")","url":"about:blank","ok":true,"status":200}
{"ts":"...","action":"navigated","url":"https://web.telegram.org/a/"}
{"ts":"...","action":"snapshot","selector":"body","interactiveOnly":false,"length":3842,"url":"https://web.telegram.org/a/"}
{"ts":"...","action":"locator","chain":"getByRole(\"link\", {\"name\":\"Log in by phone Number\"})","url":"https://web.telegram.org/a/"}
{"ts":"...","action":"click","method":"click","args":[],"chain":"getByRole(\"link\", {\"name\":\"Log in by phone Number\"}) → click()","url":"https://web.telegram.org/a/","ok":true}
{"ts":"...","action":"navigated","url":"https://web.telegram.org/a/#/login"}
{"ts":"...","action":"fill","method":"fill","args":["77054595958"],"chain":"getByLabel(\"Phone number\") → fill(\"77054595958\")","url":"https://web.telegram.org/a/#/login","ok":true}
{"ts":"...","action":"screenshot","url":"https://web.telegram.org/a/#/login"}
{"ts":"...","action":"humanClick","args":["page",100,200],"url":"https://web.telegram.org/a/#/login","ok":true}

Recording user task

Always pass the user's request via task so the log has full context:

const { page, logger } = await launchBrowser({
  task: 'Забронировать столик в Aurora на 8 марта, 19:00, 2 гостя',
  logLevel: 'verbose',
  country: 'ru',
});

Agent reasoning with logger.note()

At verbose level, the agent can record its reasoning:

logger.note('Navigating to booking page to check available slots');
await page.goto('https://restaurant.com/booking');
logger.note('Form is empty — need to fill date, time, guests before checking');

Reading logs

const { getSessionLogs, getSessionLog } = require('clawnet/scripts/browser');

// List recent sessions
const sessions = getSessionLogs();
// [{ sessionId: 'abc-123', mtime: '2026-03-01T...', size: 4096 }, ...]

// Read a specific session
const log = getSessionLog(sessions[0].sessionId);
// [{ ts: '...', action: 'task', prompt: 'Войти в Telegram...' },
//  { ts: '...', action: 'goto', method: 'goto', args: ['https://web.telegram.org'], ... },
//  { ts: '...', action: 'click', chain: 'getByRole("link") → click()', ... }, ...]

// Or from the current session
const { getSessionLog: currentLog } = await launchBrowser();
// ... do work ...
const entries = currentLog();

getCredentials()

Fetch managed proxy + CAPTCHA credentials from Clawnet API. Called automatically by launchBrowser() on fresh launch (not on reuse). Starts the 2-hour trial clock on first call. Requires CN_API_URL and agent credentials (from install, CN_AGENT_TOKEN, or CN_AGENT_ID + CN_AGENT_SECRET).

makeProxy(sessionId, country)

Build proxy config from environment variables. Supports Decodo, Bright Data, IPRoyal, NodeMaven.

Supported proxy providers

Examples

Login to a website

const { launchBrowser } = require('clawnet/scripts/browser');
const { page, snapshot } = await launchBrowser({ country: 'us', mobile: false });

await page.goto('https://github.com/login');

// Observe the page first — see what's available
const tree = await snapshot({ interactiveOnly: true });
// tree shows: textbox "Username or email address", textbox "Password", button "Sign in"

// Use semantic locators that match the snapshot
await page.getByLabel('Username or email address').fill('myuser');
await page.getByLabel('Password').fill('mypass');
await page.getByRole('button', { name: 'Sign in' }).click();

Scrape with CAPTCHA bypass

const { launchBrowser, solveCaptcha } = require('clawnet/scripts/browser');
const { page, snapshot } = await launchBrowser({ country: 'de' });

await page.goto('https://protected-site.com');

// Auto-detect and solve any CAPTCHA
try {
  await solveCaptcha(page, { verbose: true });
} catch (e) {
  console.log('No CAPTCHA found or solving failed:', e.message);
}

// Read the content area compactly
const content = await snapshot({ selector: '.content' });

Fill Shadow DOM forms

const { launchBrowser, shadowFill, shadowClickButton } = require('clawnet/scripts/browser');
const { page } = await launchBrowser();

await page.goto('https://app-with-shadow-dom.com');
await shadowFill(page, 'input[name="email"]', 'user@example.com');
await shadowClickButton(page, 'Submit');