Agent Infrastructure · Startup idea
Browser-use solves 60% of agent work. The remaining 40% — anything that touches a native app — needs a sandboxed desktop. That's a separate product, and it's still greenfield.
Why now
Claude's computer-use API and the equivalent in OpenAI's stack made desktop-grounded agents practical in late 2024. By 2026 every AI-native legal / accounting / clinical product needs a runtime. The infrastructure layer is still wide open.
The idea you could build today
Linux desktop in a VM (Firecracker or Kata Containers), pre-installed with Chrome + LibreOffice + a screen-capture loop. Expose mouse, keyboard, screenshot, OCR as MCP tools. Bill per session-minute. Record sessions for audit.
Build stack
The three repos already trying
The open source AI engineering platform for agents, LLMs, and ML models.
Framework migration
-56%
14-day velocity Δ
100 contributors
AI-Powered Photos App for the Decentralized Web. We are on a mission to protect your freedom and privacy.
Framework migration
+109%
14-day velocity Δ
100 contributors
Framework migration
+92%
14-day velocity Δ
5 contributors
Matched against the current-period startup signal panel (ai-ml, developer-tools). Rankings shift weekly as the underlying GitHub activity moves. Read the methodology.
The seed-round pattern hiding in the trendline
RPA-class repos pivoting to LLM-grounded automation in a 60-day window are the seed-round tells. Watch for repos that flip from "Selenium for desktop" framing to "computer-use runtime" framing in their README.
Most of the AI-native vertical products run against apps that have no SSH or API — desktop legal software, clinical record systems, in-house tools. The desktop is the API.
Use the signal, not just the idea
The repos above re-rank automatically as commit velocity, contributor growth, and new-repo creation move. Want the data feed for this idea wired into your own stack? The MCP server exposes every signal as a tool any agent host can query.
Updated 2026-05-18. The framing is editorial; the “three repos already trying” slot is generated from the live signal panel. Anonymity rule: we name public GitHub orgs, never individual founders or stealth teams.