Browser MCP and chrome-devtools-mcp work — until your agent clicks something with a randomized class name and the whole loop crashes. LumaBrowser is a desktop browser with an MCP server, self-healing selectors, and a token-cached page model. Plug it into Claude Desktop, Cursor, OpenClaw, or any MCP host in one config block.
LumaBrowser ships an MCP server. Add it to your claude_desktop_config.json (or your OpenClaw / Cursor MCP config), restart the host, and your agent gets browser tools.
{
"mcpServers": {
"lumabrowser": {
"command": "lumabrowser",
"args": ["--mcp"]
}
}
}Install LumaBrowser. Download from the pricing page or run npx lumabrowser start.
Edit your MCP config. On Windows: %APPDATA%\Claude\claude_desktop_config.json. On macOS / Linux: ~/.config/Claude/claude_desktop_config.json.
Restart your MCP host. Your agent now has navigate, click, fill_form, screenshot, get_source, and a dozen other tools — all backed by a real Chromium tab on your machine.
Modern frontends ship CSS-in-JS class names like css-8xk2m9 that change on every build. Most MCP browser tools delegate to Puppeteer or Playwright and crash with element not found. LumaBrowser's selector tools take an optional llmFallback — if the CSS misses, an LLM resolves the description against the live DOM and the call succeeds.
The frontend pushes a new build at 4pm. Class names change. Your agent loop crashes on the first click() and you find out from an alert at 8pm.
click({ selector: ".product-buy-btn" })
→ Error: element not foundPass an llmFallback alongside your selector. When the CSS misses, an LLM looks at the live DOM, picks the right element by description, and the call returns.
click({
selector: ".product-buy-btn",
llmFallback: "the primary buy button
under the price"
}) → okIf your agent is reading entire DOM trees to figure out where to click, you're burning tens of thousands of tokens per turn. LumaBrowser's Template Builder runs once per domain, generates a generalized selector map, and caches it — so subsequent visits return a tiny structured map instead of the whole page.
The first time your agent hits a domain, the Template Builder analyzes the page and writes a generalized selector map to local SQLite. Every subsequent visit on that domain pulls the cached map — same structural output, ~30× fewer tokens. Multiply by every page your agent loops over and the math gets very different.
A focused set of MCP tools, not a 60-tool kitchen sink. Every interaction tool that targets the DOM accepts an llmFallback description.
Open tabs, route them to URLs, wait on a selector or page state, close them when done. The agent gets a real tab list it can reason about, not a single sandboxed page.
Click, type, scroll, send keystrokes, run JS in page context. Selector-based tools fall back to an LLM description if the CSS misses, so a frontend redesign doesn't end the run.
Grab cleaned page text, structured element properties, full PNG screenshots, or the cached generalized template. Three formats means the agent picks the cheapest one for the job.
There are a lot of ways to give an agent a browser right now. Here's how the local-desktop-app approach compares to the alternatives engineers usually try first.
| Browser MCP / Claude in Chrome | chrome-devtools-mcp | Browserbase | Browserless | LumaBrowser | |
|---|---|---|---|---|---|
| Where it runs | Chrome / Edge extension | Local Chrome via Puppeteer | Vendor cloud only | Vendor cloud or self-host | Local desktop app |
| MCP host support | Claude Desktop / Code only | Any MCP host | Custom SDK + MCP server | HTTP / WS API, MCP via wrappers | Any MCP host |
| Pricing model | Free, gated to Anthropic | Free, OSS | Free 1 hr, then $20+/mo | Free 1k units, then $25+/mo | Free, no caps |
| Self-healing selectors | No | No (raw Puppeteer) | Stagehand AI actions | No | LLM fallback per call |
| Token-cached page maps | No | No | No | No | Template Builder |
| Local model support | No (Anthropic only) | Bring your own host | Cloud-only | Cloud-only | LM Studio, WebGPU, OpenAI, Anthropic |
| Where the data lives | Anthropic + your Chrome | Your machine | Their cloud | Their cloud or yours | Your machine |
Pricing as listed by each vendor at time of publication. Cloud headless-browser pricing is metered — Browserbase by browser-hours, Browserless by 30-second “Units” — so cost scales with how often your agent loops. Higher tiers (Browserbase Startup at $99/mo, Browserless Starter at $140/mo annually) raise the included quota and concurrency.
This comparison reflects publicly available pricing and feature information gathered to the best of our knowledge from each vendor's public materials. Vendors update plans frequently and we're a small team — if anything here looks wrong, please email [email protected] with the correction and a source, and we'll update the page.
Bring your own MCP host. Bring your own model.
~/.cursor/mcp.jsonconnectOverCDPhttp://localhost:1234/v1; fully local inferenceLumaBrowser is the local desktop browser your agent already wishes you had wired up. Free to download, plugs into your existing MCP host, runs with the model you're already paying for — or no model at all if you want to drive it from a script.