Agent Infrastructure · Startup idea
Computer-use agents: the runtime for full-desktop automation
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
- ·Vercel Sandbox or self-hosted Firecracker
- ·noVNC / WebRTC for live view
- ·Tesseract / Claude vision for OCR
- ·Postgres + S3 for the session recording
The three repos already trying
Pulled live from our current-period signal index.
No repos in the current period have crossed our minimum-volume + sector match for this idea. That can mean two things, the niche is too new for public GitHub signal to show up yet, or the buildable wedge is genuinely greenfield. Either way, watch this page; matches refresh weekly.
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.
Frequently asked
Why not just give the agent SSH?+
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
Watch this idea live, every week.
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.
Related ideas
Updated 2026-08-25. 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.