Browser Use with CDP

Drive the browser through its native Chrome DevTools Protocol over the remote-debugging WebSocket. This works with zero dependencies: launch the browser with --remote-debugging-port, then talk JSON over fetch and the built-in WebSocket global (available in Bun and Node ≥ 22 — no ws package).

If the project already has Playwright or Puppeteer installed, using it is usually simpler — reach for raw CDP when no automation library is available, when protocol-level control is needed, or when recording a deterministic visual demo.

Protocol reference: https://chromedevtools.github.io/devtools-protocol/. The running browser's exact schema is at http://127.0.0.1:<port>/json/protocol; tip-of-tree docs can differ from the installed version.

Managed cloud sandbox default: visible browser

When running in a cloud sandbox, default every browser task to the visible managed desktop, even when the user did not explicitly ask to watch. Most browser tasks exist because plain HTTP is not enough; a headless browser is more likely to trigger bot protection and gives the user no way to observe or take over. This matters especially for clicking or typing, forms, sign-in, checkout/payment, CAPTCHAs or bot protection, and user handoff.

For the first managed-sandbox browser window, use the skill's launcher instead of assembling Chrome, DISPLAY, or Xvfb commands yourself:

/root/.letta/cloud-skills/browser-use/scripts/open-visible-browser.sh 'https://example.com'

It starts the managed desktop and launches Chrome through the persistent Cua Driver with its required root flag. Then load computer-use for visible interaction. If protocol-level control is necessary, use Cua Driver's explicit browser_prepare flow after binding the exact visible window; do not pass remote-debugging flags through launch_app. The launcher exits zero only after Cua Driver reports an on-screen browser window. If it exits nonzero, stop and report the launch failure instead of claiming the browser opened.

When the user asks to review, watch, or take over, leave that browser window open after the task. Do not kill or close it before replying.

1. Run start-letta-desktop and use its exit status as the result. Warnings from optional services do not mean startup failed when the command exits 0. If it exits nonzero, stop and report that the managed desktop is unavailable. Never create another Xvfb, VNC server, or private display: the Computer viewer only shows the managed desktop.
2. Load computer-use, inspect the managed desktop, and use Cua Driver to operate an existing Chrome window or launch Chrome there. When launching, round-trip Chrome's launch_path from `cua-driver call list_apps '{}' instead of rebuilding it; the managed launch path carries required flags. For forms, sign-in, checkout, CAPTCHA, and bot-protected pages, keep using Cua Driver so the interaction remains visible and available for user takeover.
3. Use CDP only when protocol-level inspection or deterministic automation is needed. Bind the exact visible browser window with Cua Driver, then use its explicit browser_prepare flow. Do not pass remote-debugging flags through launch_app. Include --no-sandbox when running Chrome as root. Do not add --headless or override DISPLAY.
4. Use headless mode only for work the user explicitly wants in the background and that cannot require interaction or handoff, such as read-only scraping, CI, or screenshot/PDF generation.
5. Verify the result through the managed desktop window (Cua Driver window state or screenshot), not only through DOM output or a screenshot from a separate process.

A headless page does not satisfy a request to open or reopen a site in the user-visible browser.

Workflow

1. Find a Chromium-based browser (below). If none exists, see "No Chrome installed".
2. In a managed cloud sandbox, follow the visible-browser default above. Otherwise, launch with a dedicated profile and remote debugging. Never attach to the user's normal profile unless explicitly asked.
3. Discover targets via /json/list; pick the "page" target by URL or title.
4. Connect to its webSocketDebuggerUrl and enable only the domains you need (usually Page, Runtime, DOM, Input; add Network, Log when debugging).
5. Inspect before acting: find elements by accessible name, label, text, role, stable ID, or placeholder — not generated classes or child indexes.
6. Act through Input.* for user-like interactions; use Runtime.evaluate for inspection, coordinate math, and setup with no meaningful user interaction.
7. Wait on observable state, never fixed sleeps alone.
8. Verify the result (DOM state, URL, screenshot, console/network events).
9. Clean up temporary background work: stop screencasts and close the WebSocket. Kill a browser you launched only when the user did not ask to review, watch, or take over the visible window.

Finding the browser

Any Chromium-based browser supports CDP (Chrome, Chromium, Edge, Brave). Probe in order:

# macOS
for c in "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
         "/Applications/Chromium.app/Contents/MacOS/Chromium" \
         "/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge" \
         "/Applications/Brave Browser.app/Contents/MacOS/Brave Browser"; do
  [ -x "$c" ] && { echo "$c"; break; }
done

# Linux
for c in google-chrome google-chrome-stable chromium chromium-browser microsoft-edge brave-browser; do
  command -v "$c" && break
done

On Windows, check %ProgramFiles%\Google\Chrome\Application\chrome.exe, %ProgramFiles(x86)%\..., %LocalAppData%\Google\Chrome\Application\chrome.exe, and the same patterns for Microsoft\Edge.

No Chrome installed

Any browser found by the probe above works identically — use it. If truly no Chromium-based browser exists, do not install or download one automatically. Tell the user that browser use requires Chrome or another Chromium-based browser and recommend either:

1. Install Chrome on the current computer, then retry the browser task.
2. Teleport the conversation back to its Cloud sandbox, where the managed browser is already installed.

Wait for the user to choose. Do not silently replace the browser task with plain HTTP or claim browser automation succeeded.

Launching

Outside a managed cloud sandbox, use a disposable profile and a fixed port. Chrome refuses to run as root without --no-sandbox, so add that flag when id -u is 0:

chrome_args=( \
  --remote-debugging-port=9222 \
  --user-data-dir=/tmp/cdp-profile \
  --window-size=1440,900 \
  --force-device-scale-factor=1 \
  --no-first-run \
  --no-default-browser-check \
)
[ "$(id -u)" -eq 0 ] && chrome_args+=(--no-sandbox)
"$CHROME" "${chrome_args[@]}" https://example.com

Outside a managed cloud sandbox, add --headless=new only for explicitly invisible work or when no display exists. In a managed cloud sandbox, follow the visible-browser rule above and never replace its managed display with a private one. With --remote-debugging-port=0, read the chosen port from <user-data-dir>/DevToolsActivePort. Launch in the background and poll http://127.0.0.1:9222/json/version until it responds.

HTTP endpoints: /json/version (browser metadata + browser-level WebSocket URL), /json/list (targets), PUT /json/new?<url> (open tab), /json/activate/<id>, /json/close/<id>, /json/protocol (schema).

Attach to the page target for Page/DOM/Runtime/Input work; use the browser target only for browser-wide commands (target control, downloads, browser contexts).

Minimal CDP client

CDP messages are JSON with monotonically increasing request ids. Run with bun:

const targets = await (await fetch("http://127.0.0.1:9222/json/list")).json();
const target = targets.find((t: any) => t.type === "page");
if (!target) throw new Error("No page target");

const ws = new WebSocket(target.webSocketDebuggerUrl);
await new Promise((resolve, reject) => {
  ws.onopen = resolve;
  ws.onerror = reject;
});

let nextId = 0;
const pending = new Map();
ws.onmessage = (event) => {
  const msg = JSON.parse(String(event.data));
  if (!msg.id) return handleEvent(msg); // Page.loadEventFired, Log.entryAdded, ...
  const p = pending.get(msg.id);
  if (!p) return;
  pending.delete(msg.id);
  msg.error ? p.reject(new Error(JSON.stringify(msg.error))) : p.resolve(msg.result);
};
ws.onclose = () => {
  for (const p of pending.values()) p.reject(new Error("socket closed"));
  pending.clear();
};

function send(method: string, params: object = {}): Promise<any> {
  return new Promise((resolve, reject) => {
    const id = ++nextId;
    pending.set(id, { resolve, reject });
    ws.send(JSON.stringify({ id, method, params }));
  });
}

await send("Page.enable");
await send("Runtime.enable");

Navigations destroy execution contexts and can invalidate in-flight Runtime.evaluate calls; retry after observing the new document.

Inspecting the page

Start with a concise UI inventory:

const result = await send("Runtime.evaluate", {
  expression: `JSON.stringify({
    buttons: [...document.querySelectorAll('button')].map((el) => ({
      text: el.innerText.trim(), aria: el.getAttribute('aria-label'), title: el.title
    })).filter((x) => x.text || x.aria || x.title),
    inputs: [...document.querySelectorAll('input, textarea, [contenteditable=true]')].map((el) => ({
      tag: el.tagName, type: el.type, placeholder: el.placeholder,
      aria: el.getAttribute('aria-label'), value: el.value
    }))
  })`,
  returnByValue: true,
});

Use awaitPromise: true for async expressions and userGesture: true when the page requires user activation. Treat exceptionDetails in the result as an error even though the CDP command itself succeeded.

document.querySelector does not cross shadow boundaries — traverse open shadow roots explicitly; for closed shadow roots or remote-object work use the DOM domain (DOM.getDocument, DOM.querySelector, DOM.getBoxModel).

Clicking and typing

Compute coordinates in CSS pixels immediately before acting, then send native input events:

async function point(expr: string) {
  const r = await send("Runtime.evaluate", {
    expression: `(() => {
      const el = ${expr};
      if (!el) return null;
      el.scrollIntoView({ block: 'center', inline: 'center' });
      const b = el.getBoundingClientRect();
      return { x: b.left + b.width / 2, y: b.top + b.height / 2 };
    })()`,
    returnByValue: true,
  });
  if (!r.result.value) throw new Error(`Element not found: ${expr}`);
  return r.result.value;
}

async function click(expr: string) {
  const { x, y } = await point(expr);
  await send("Input.dispatchMouseEvent", { type: "mouseMoved", x, y });
  await send("Input.dispatchMouseEvent", { type: "mousePressed", x, y, button: "left", buttons: 1, clickCount: 1 });
  await send("Input.dispatchMouseEvent", { type: "mouseReleased", x, y, button: "left", buttons: 0, clickCount: 1 });
}

Focus an editable element (click it), then insert text:

await click(`document.querySelector('input[aria-label="Search"]')`);
await send("Input.insertText", { text: "search terms" });

Use Input.insertText for text and Unicode; use paired Input.dispatchKeyEvent (keyDown + keyUp with key, code, windowsVirtualKeyCode) for Enter, Escape, arrows, Tab, and shortcuts. Modifier bits: Alt=1, Ctrl=2, Meta=4, Shift=8.

Native <select> and framework-controlled inputs may need the prototype setter plus bubbling events:

const setter = Object.getOwnPropertyDescriptor(HTMLInputElement.prototype, "value").set;
setter.call(input, "new value");
input.dispatchEvent(new Event("input", { bubbles: true }));
input.dispatchEvent(new Event("change", { bubbles: true }));

Prefer real Input.* events for the behavior being demonstrated or tested; direct DOM mutation is fine for deterministic setup and inspection.

Waiting reliably

A returned command does not mean the action completed. Poll the state that proves completion:

async function waitFor(expr: string, timeoutMs = 30_000) {
  const start = Date.now();
  while (Date.now() - start < timeoutMs) {
    const r = await send("Runtime.evaluate", { expression: `Boolean(${expr})`, returnByValue: true });
    if (r.result.value) return;
    await new Promise((res) => setTimeout(res, 250));
  }
  throw new Error(`Timed out waiting for ${expr}`);
}

await waitFor(`document.body.innerText.includes('Saved')`);

For navigation, wait on Page.loadEventFired or a lifecycle networkIdle event — but SPA route changes may emit neither, so prefer the UI condition that actually matters.

Screenshots, PDF, and video

const shot = await send("Page.captureScreenshot", { format: "png" });
await Bun.write("screenshot.png", Buffer.from(shot.data, "base64"));

captureBeyondViewport: true for full-page; Page.getLayoutMetrics + clip for exact regions; Page.printToPDF for PDFs. For video recording with Page.startScreencast and demo-polish tips, read references/recording.md.

Debugging failures

Enable Network and Log, then watch Network.requestWillBeSent, Network.responseReceived, Network.loadingFailed, Runtime.consoleAPICalled, Runtime.exceptionThrown, and Log.entryAdded. Fetch bodies with Network.getResponseBody. Never log authorization headers, cookies, API keys, passwords, or response bodies containing secrets — scrub before returning output to context.

Common failure modes:

• ECONNREFUSED on the port: browser exited, wrong port, or debugging not enabled — check /json/version first.
• No matching target: inspect /json/list; match by URL/title, don't take the first page blindly.
• Execution context destroyed: the page navigated; wait for the new document and re-evaluate.
• Click misses an existing element: scroll into view and recalculate the box immediately before dispatching.
• Typed text doesn't stick in a controlled input: focus + Input.insertText, or the prototype-setter pattern above.
• Opening DevTools disconnects automation: embedded DevTools can detach other clients — don't open DevTools during a run.
• Page commands fail on the browser endpoint: attach to the page target.

Safety

Browser automation acts with the user's browser authority.

• Use a disposable profile by default.
• Do not submit purchases, publish content, send messages, delete data, or accept consequential dialogs without explicit authorization.
• Do not extract saved passwords, tokens, cookies, or unrelated browsing data.
• Stay within the requested origin and workflow.
• Keep the remote-debugging listener on loopback (127.0.0.1) unless the user explicitly needs remote access and has authentication in place.