LumaBrowser is a programmable browser for AI agents, QA engineers, and backend developers. It ships with a REST API, a built-in MCP server, and CDP-based network interception, and every click or find can fall back to an LLM when selectors break. Install in one command; bring your own local, OpenAI, or Anthropic model.
# Create a tab and navigate curl -X POST http://localhost:3000/api/browser/tabs \ -H "Content-Type: application/json" \ -d '{"url": "https://example.com"}' # Click with LLM fallback - never break on selector changes curl -X POST http://localhost:3000/api/browser/tabs/0/click \ -d '{"selector": "button.submit", "llmFallback": "Click submit"}' # Get the page content curl http://localhost:3000/api/browser/tabs/0/source?type=text
npx lumabrowser start
LumaBrowser in Action, and it can do... so much more.
LumaBrowser exposes a built-in Model Context Protocol (MCP) server, so Claude Desktop, OpenClaw, and any MCP-compatible agent can navigate pages, click elements, fill forms, and read the DOM natively. No glue code, no custom tool wrappers; the browser shows up as first-class tools to whatever LLM you point at it.
LumaBrowser's selector resolver tries stable attributes first, data-testid, aria-label, placeholder, visible text, with zero LLM cost, then escalates to a model grounded in a live DOM snapshot when nothing matches. Every resolved selector is validated before it fires, and the template cache warms itself on every run.
LumaBrowser's Network Watcher uses the Chrome DevTools Protocol to capture HTTP traffic from inside the browser process, no root certificates, no MITM plumbing. Define URL patterns, pick methods, and the full request and response are forwarded to any backend webhook you configure.
Three lines to navigate, click, and extract, against any website.
# Create a tab and navigate to a page
curl -X POST http://localhost:3000/api/browser/tabs \
-H "Content-Type: application/json" \
-d '{"url": "https://news.ycombinator.com"}'
# Click with LLM fallback - if the selector breaks, AI finds the right element
curl -X POST http://localhost:3000/api/browser/tabs/0/click \
-H "Content-Type: application/json" \
-d '{"selector": ".titleline > a", "llmFallback": "Click the first story link"}'
# Get the page content as clean text
curl http://localhost:3000/api/browser/tabs/0/source?type=textimport requests
# Create a tab and navigate
requests.post("http://localhost:3000/api/browser/tabs", json={
"url": "https://news.ycombinator.com"
})
# Click with LLM fallback - resilient to DOM changes
requests.post("http://localhost:3000/api/browser/tabs/0/click", json={
"selector": ".titleline > a",
"llmFallback": "Click the first story link"
})
# Fill a form - each field has its own LLM fallback
requests.post("http://localhost:3000/api/browser/tabs/0/fill", json={
"fields": [
{"selector": "input[name=q]", "value": "LumaBrowser",
"llmFallback": "The search input"}
]
})// Create a tab and navigate
await fetch("http://localhost:3000/api/browser/tabs", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({ url: "https://news.ycombinator.com" })
});
// Click with LLM fallback - AI resolves the selector if it breaks
await fetch("http://localhost:3000/api/browser/tabs/0/click", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
selector: "button.submit",
llmFallback: "Click submit"
})
});
// Wait for an element - LLM resolves if selector is stale
await fetch("http://localhost:3000/api/browser/tabs/0/wait", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
selector: ".success-msg",
llmFallback: "The success confirmation message"
})
});Real request/response flows showing what LumaBrowser handles for you.
{ "url": "https://acme.com/signup" }{ "id": 0, "url": "https://acme.com/signup", "title": "Sign Up - Acme" }{
"fields": [
{ "selector": "#full_name", "value": "Jane Doe",
"llmFallback": "The full name input field" }
],
"llmFallback": "Fill the signup form"
}#full_name not found; site renamed it to data-field-xk92llmFallback is set → triggering AI resolution...
{
"pageInfo": { "selectorStability": "unstable", "jsRendered": true },
"elements": [
{ "Name": "Full Name Input", "CssSelector": "input[data-field-xk92]",
"FallbackSelector": "form .field-group:first-child input",
"Type": "Input" },
{ "Name": "Email Input", "CssSelector": "input[data-field-ml47]",
"Type": "Input" },
{ "Name": "Submit Button", "CssSelector": "button.btn-register",
"Type": "Button" }
]
}input[data-field-xk92]Prompt: "Action: fill | Description: The full name input field
Page template elements:
- "Full Name Input" (Input): input[data-field-xk92]
- "Email Input" (Input): input[data-field-ml47]
..."
→ input[data-field-xk92]{
"success": true,
"message": "Form filled successfully (resolved by LLM fallback)",
"data": {
"filled": 1,
"resolvedSelector": "input[data-field-xk92]"
}
}Compare: ?type=full vs ?type=clean vs ?type=text
<!DOCTYPE html> <html lang="en" data-theme="dark" data-build="a8f3c"> <head> <meta charset="UTF-8"> <meta name="viewport" content="..."> <title>Dashboard - Acme Corp</title> <link rel="stylesheet" href="/assets/main.c7x92k.css"> <link rel="stylesheet" href="/assets/vendor.d83hf.css"> <script src="/assets/runtime.k29d.js" defer></script> <script src="/assets/app.m48fk.js" defer></script> <style>.css-1a2b3c{display:flex} .css-4d5e6f{margin:0 auto;max-width:1200px} .css-7g8h9i{font-family:Inter,sans-serif} /* ... 847 more lines of CSS ... */</style> <script>window.__NEXT_DATA__={props:{pageProps: {session:{user:{id:"usr_382",name:"Jane"}, token:"eyJhbG..."}}},page:"/dashboard"} </script> </head> <body> <div id="__next"> <div class="css-1a2b3c"> <nav class="css-k3m8x2"> <img src="data:image/svg+xml;base64,PHN2..." alt="" class="css-p9q2r1" /> <a href="/" class="css-m4n5o6">Dashboard</a> <a href="/settings" class="css-m4n5o6">Settings</a> </nav> <main class="css-4d5e6f"> <h1 class="css-7g8h9i">Welcome back, Jane</h1> <div class="css-a1b2c3"> <div class="css-d4e5f6">Revenue</div> <span class="css-g7h8i9">$12,450</span> </div> <!-- ... 200+ more nodes ... -->
<nav>
<a href="/">Dashboard</a>
<a href="/settings">Settings</a>
</nav>
<main>
<h1>Welcome back, Jane</h1>
<div class="metric-card">
<div>Revenue</div>
<span>$12,450</span>
</div>
<div class="metric-card">
<div>Active Users</div>
<span>1,284</span>
</div>
<table class="data-table">
<thead>
<tr><th>Product</th><th>Sales</th><th>Revenue</th></tr>
</thead>
<tbody>
<tr><td>Widget Pro</td><td>342</td><td>$8,550</td></tr>
<tr><td>Gadget X</td><td>156</td><td>$3,900</td></tr>
</tbody>
</table>
<a href="/reports">View full report →</a>
</main>
Dashboard | Settings
Welcome back, Jane
Revenue: $12,450
Active Users: 1,284
Product Sales Revenue
Widget Pro 342 $8,550
Gadget X 156 $3,900
View full report →
{
"prompt": "Log into acme.com/login with [email protected] / demo123,
go to settings, change company name to 'Acme Industries'",
"autoCloseTab": false,
"includeScreenshot": true
}1. navigate → acme.com/login (template attached) 2. fill_form → email + password 3. click → Login button (url changed → /dashboard) 4. wait_for → .dashboard-content ✓ 5. navigate → /settings (template generated) 6. fill_form → company name = "Acme Industries" 7. click → Save Changes ✓
{
"success": true,
"summary": "Logged in and updated company name to Acme Industries",
"finalResponse": "Done! I logged into acme.com, navigated to...",
"tabId": 0,
"iterations": 4,
"durationMs": 18420,
"toolCalls": [
{ "tool": "navigate", "durationMs": 1200 },
{ "tool": "fill_form", "durationMs": 340 },
{ "tool": "click", "durationMs": 890 },
{ "tool": "wait_for", "durationMs": 1100 },
{ "tool": "navigate", "durationMs": 1450 },
{ "tool": "fill_form", "durationMs": 280 },
{ "tool": "click", "durationMs": 760 }
],
"screenshot": "data:image/png;base64,iVBOR..."
}What the AI Chat agent actually does behind the scenes, with Template Builder enabled for faster targeting.
{ "prompt": "Log into acme.com/login with [email protected] / demo123, go to settings, change company name to 'Acme Industries'." }https://acme.com/logininput#email, input#password, button.login-btninput#email = "[email protected]", input#password = "demo123", selectors from template, instant matchbutton.login-btn, "Log In" button from template.dashboard-content to appear → visible in 1.2s/settings, template auto-generated (no cache), 8 elements extractedinput#company-name, button.save-settings, etc.input#company-name = "Acme Industries", from freshly generated templatebutton.save-settings, "Save Changes".toast-success → "Settings saved successfully" appears in 0.8sA built-in LLM assistant that can see and control the browser. Ask it to navigate pages, fill forms, extract data, or run multi-step automations, conversationally. It has full access to every browser tool and streams responses in real time.
Here are the top 3 stories on Hacker News right now:
AI Chat doesn't just reply with text; it builds artifacts that render live next to the conversation: code, web pages, SVG, Markdown, generated images, and fully interactive mini-apps.
Here's an interactive tip calculator; drag the sliders and it recalculates instantly:
Instead of one do-everything assistant, define focused sub-agents, each with its own system prompt, its own curated set of tools, and optionally its own model. The main chat can delegate to them, or you can call them over the REST API.
list_agents / chat_with_agent/api/ext/agent-manager (SSE or JSON)Delegating to Research Agent (web-search + read-only tools, pinned to a fast local model)…
It found 5 sources and returned a 3-bullet brief.
A bundled image server built on stable-diffusion.cpp generates and edits images entirely on your machine, no cloud, no per-image fee, nothing leaves your device.
generate_image / edit_image return artifactsGenerated locally on your GPU and refined with the edit model, returned as an image artifact you can open or download.
An AI build agent that turns a plain-English request into a real, installable LumaBrowser extension. It plans the files, writes them with live syntax checking, then installs and activates the extension, no restart.
Wrote manifest.js and main.js (both validated ✓), installed and activated. Try it on any article.
Monitor HTTP traffic at the browser level using Chrome DevTools Protocol. Define URL patterns, and LumaBrowser captures matching requests and forwards the full request/response to your webhook, no proxy, no root certs.
The first time you visit a page, the LLM analyzes the full DOM and generates a cached selector map. Every subsequent visit uses the cached template, slashing token usage by 95%+ and making automation near-instant.
generate, validate, repair, check any selector against a live tab in milliseconds, regenerate only when it breaks{
"elements": [
{ "name": "Login Button",
"primary": "button[data-testid='login']",
"fallbacks": ["button.auth-cta", "#login-btn"] },
{ "name": "Email Field",
"primary": "input[name='email']",
"fallbacks": ["input[type='email']"] }
]
}
Schedule recurring LLM-driven automations. Each execution runs the same headless AgentRunner that powers AI Chat, with the full template workflow, structured JSON output, and webhook callbacks. Set it and forget it.
{"rate":""} and the agent fills it exactlyWatch any web page, or any specific element on it, for content changes. Pick what you care about with a click, and get a webhook the second it updates. Real browser, so JS-rendered pages and SPAs just work. Fully local. Full details →
+/− summary on every change/api/page-change-detector, CRUD, history, manual check, picker{
"monitorName": "Acme Pricing Page",
"url": "https://acme.com/pricing",
"timestamp": "2026-05-03T17:22:08Z",
"diffSummary": "+2 / -1\n+ $19.99 - In stock\n- $24.99 - Out of stock",
"changeCount": 17
}
A first-class blocker runs at the Electron session layer, independent of browser extensions. Because Electron doesn't implement chrome.webRequest, Chrome-based ad blockers like uBlock Origin cannot actually block anything inside LumaBrowser. This is the replacement; and it works on every tab, every session, including those spun up by automation.
@ghostery/adblocker-electron (MPL-2.0)Load unpacked Chrome extensions directly into every tab session. Good for content-script extensions like Dark Reader, Stylus, or password manager autofill; but read the limitations carefully, because Electron only implements a subset of the Chrome extension APIs.
persist:main-* partitions)storage · tabs · runtime · i18n · webNavigation · managementNote: webRequest and declarativeNetRequest are not implemented, so Chrome-based ad blockers will load but will not block anything. Use the built-in Native Ad Blocker instead.
Captures browser Notification API calls from any page and forwards them to your webhook. Monitor alerts, messages, or status changes from any web app, without polling.
{
"source": "app.example.com",
"title": "New message from Sarah",
"body": "Hey, the deploy finished!",
"timestamp": "2026-03-29T14:22:00Z",
"url": "https://app.example.com/chat"
}
Local models without the setup tax. LumaBrowser reads your hardware, recommends a curated GGUF model that actually fits, and resumably downloads it onto a bundled llama.cpp server, one click, no API key, nothing leaves your machine. Then it's a streaming chat with live artifacts.
Every feature is a module on one small runtime of six shared services. Most extensions plug straight into the Core; a few build on each other; and yours drops in anywhere. Hover an extension to see which services it taps — or hover a service to see who uses it.
A lightweight JavaScript data grid with virtual scrolling, sortable columns, context menus, column reordering, and CSV export. Fast and dependency-free.
A lightweight reactive UI library for building dynamic interfaces with minimal overhead. Simple API, no build step, no virtual DOM.
The textbook price comparison engine. Find the best deals on textbooks across multiple retailers in one search.