This Is Reduck — Browser Infrastructure for AI Agents
We could have spent weeks perfecting this first blog post. But at Reduck, we prefer to get straight to the point.
We’re a new French startup with three repeat founders—Daniel (CEO), Louis (CTO), and Raphaël (COO)—and we’ve just raised over €2.8 million to solve a problem that’s been bothering us for a while: browser automation is fundamentally broken for AI agents.
Who We Are
Daniel and Raphaël previously built Mithril Security together, where they collaborated with OpenAI on privacy-preserving AI infrastructure and launched LaVague—the first open-source AI web agent in 2024, which hit 6,000 GitHub stars. They sold the company to H Company, a startup founded by DeepMind engineers that raised $220M.
Louis is a world-class engineer who co-founded Secrecy (end-to-end encryption platform with patented cryptographic protocols), collaborated with Stanford on biomedical ML research, and led his team to France’s best-ever finish at the ICPC 2020 World Finals—the world’s top competitive programming contest.
When Daniel and Raphaël saw the gap between AI agent hype and reality during the LaVague days, they knew exactly who to bring on board to fix it.
The Problem
AI agents have reached an inflection point. The models are good enough. The demand is real. But they keep breaking in production.
Not because the LLMs aren’t smart enough—they are. They break because browsers weren’t designed to be piloted by AI, and the infrastructure to bridge that gap doesn’t exist.
Agents fail against anti-bot protections. They lose authentication mid-workflow. They can’t handle multi-tab orchestration. They re-run expensive operations because they can’t persist state between sessions.
And here’s the bigger issue: most services don’t have APIs. The ones that do often don’t cover the workflows you actually need. How do you programmatically fetch invoices from Anthropic? Or automate supplier portal logins across dozens of different systems? You can’t. Not reliably.
So companies are stuck between overhyped agent solutions that don’t work and building everything themselves from scratch. Neither is sustainable.
What We’re Building: Reduck
Reduck is open browser infrastructure for AI agents. It handles what breaks most automation: anti-bot protections, authenticated sessions, multi-tab orchestration, and state persistence — so your agent doesn’t have to.
Today, Reduck ships as a CLI tool that any AI agent can call. Whether you’re running an OpenClaw assistant, a NanoClaw container, Claude Code, or your own custom agent — your agent calls reduck from the terminal and gets reliable browser execution back. Your agent thinks the logic; Reduck runs it in the browser.
The architecture has two layers: an infra layer (open source core, browser automation stack routed through Reduck’s servers) and an API layer (agentic orchestration on top, free tier with rate limits, paid for scale).
Our First Product
On top of that infrastructure, we’re launching a Chrome Extension that puts an AI agent directly in your browser. You tell it what you need — it does it. No scripts, no setup, no manual clicking. Just describe the task and reduck does it.
Because it’s connected to your actual browser, it uses your existing sessions — your cookies, your authentication, your logged-in accounts. No credentials to manage, no re-authentication. It operates as you, on the web you already have access to. And since it runs inside a real browser with a real user profile, most anti-bot protections simply don’t trigger — the traffic looks exactly like you browsing normally.
What’s Next
We’re building deliberately. Too many AI agent companies overpromised and underdelivered. That’s why you won’t find much about us online yet — we’re staying quiet until we have something real to show.
The roadmap: an MCP server so any LLM-powered agent (Claude, ChatGPT, or others) can use Reduck as a native tool; a full SDK for deeper programmatic integration.
Our long-term vision: build a semantic index of the web that turns UIs into reliable, robust APIs — building blocks for any automation you can imagine.
Why This Matters
The web wasn’t designed to be automated by AI. But with the right infrastructure, it can be.
We’re not trying to build another demo-ware agent that looks cool in a tweet but breaks in production. We’re building foundational infrastructure—the boring, unsexy layer that makes everything else actually work.
If you’re building AI agents and frustrated by reliability issues, or if you’re just curious about what we’re doing, reach out. We’d love to hear from you.
Reduck is based in France. We’re currently in stealth mode while we build. Stay tuned.