How to Build a Manus-Style AI Employee With Claude, OpenClaw, Lightpanda and Mem0

#ai #agents #opensource #tutorial

Meta closed its acquisition of Manus in December 2025 for a reported price above $2 billion (TechRadar). Manus sold AI agents that each get their own virtual computer, and it reached about $100M in annualized revenue roughly eight months after its March 2025 launch (Trending Topics). It did not train its own model.

That last fact is the interesting one for developers. The value was in the harness: computer, browser, memory, permissions. Most of that harness now exists as open source. This post walks through how the pieces fit if you want to build a narrow "AI employee" for one type of small business.

The four layers

Layer	Job	Open source option	Notes
Brain	Plan steps, choose tools, check results	Claude via API (paid per use)	The only paid part of the stack
Hands	Files, shell, scripts, chat apps, cron	OpenClaw	Self-hosted, model agnostic, can run sandboxed
Browser	Log in, fill forms, read pages	Lightpanda	Zig headless browser, speaks CDP, AGPL-3.0, beta
Memory	Remember clients, rules, preferences	Mem0	Memory layer with an API

OpenClaw had about 392K GitHub stars and Mem0 about 67K by mid October 2026 according to this DEV roundup. Lightpanda's own benchmark claims about 11x faster execution and 9x less memory than headless Chrome. Treat that as a vendor number and measure on your own workload.

Architecture

Owner (WhatsApp / Slack / web dashboard)
        |
   Task intake  ->  Task queue (per customer)
                        |
               Agent worker (sandbox per customer)
               +----------------------------------+
               | OpenClaw runtime                 |
               |   brain: Claude API              |
               |   tools: files, scripts, email   |
               |   browser: Lightpanda over CDP   |
               |   memory: Mem0 (customer scope)  |
               +----------------------------------+
                        |
     Policy check (approve pay / delete / new recipient)
                        |
        Action log + screenshots -> owner's morning report

Step by step

1. Pick one job with a clear finish line. "Send yesterday's unpaid invoice reminders from the billing portal" is a good first job. "Run my business" is not.

2. Give each customer an isolated worker. One container or VM per customer, with its own browser profile and credentials vault. Never share browser sessions across customers.

3. Point the browser at Lightpanda. Because it speaks the Chrome DevTools Protocol, existing Puppeteer or Playwright scripts can connect to it instead of launching Chrome. Keep a Chrome fallback for sites that Lightpanda's beta does not render correctly yet.

// Playwright connecting to a CDP endpoint (Lightpanda or Chrome)
import { chromium } from "playwright";
const browser = await chromium.connectOverCDP(process.env.CDP_URL);
const page = await browser.newPage();
await page.goto("https://billing.example.com/invoices?status=unpaid");

4. Scope memory per customer. Store facts like "invoices go out Fridays" or "never book before 9 a.m." in Mem0 under a customer ID, and load them into the system prompt for each run.

5. Put a policy layer between the brain and the world. The model proposes actions. Deterministic code decides whether they run, ask for approval, or are blocked. Payments, deletions and messages to new recipients always ask.

6. Log everything. Every tool call, URL, screenshot and model response with token cost. The morning report the owner reads is generated from this log, and so is your eval set.

7. Cap spend per customer. Set a monthly model budget in code. When a customer hits it, the agent pauses and notifies you instead of silently burning tokens.

Where these projects usually break

Login and 2FA. Plan a human handoff for one-time codes from day one.
Sites that change. Prefer stable selectors and text over coordinates, and alert on repeated failures.
Runaway loops. Limit steps per task and stop on repeated identical actions.
Licensing. Lightpanda is AGPL-3.0. Check what that means for how you deploy it before you sell a hosted product.

Pricing it

A niche agent at $99 a month across 300 small offices would be $29,700 a month. That is a goal to plan against, and your gross margin depends almost entirely on model calls and how many browser sessions you can pack per server. That is why the memory footprint of the browser matters more than it looks.

At Axionry, an AI product development studio, we build agent products like this, with sandboxing, approvals and audit logs, from demo to launch in a few weeks. If you are building one, book a call.

What job would you give an AI employee first? Tell me in the comments.

DEV Community