AgentDesk is a control room for AI agents, built to be remixed as a real SaaS starter — not a single-screen chatbot demo.
Every run goes through one clean interface, AgentRuntime, with two sides: a fully deterministic demo runtime that needs zero API keys and works instantly, and a provider runtime that makes a real request to whichever model provider you've configured. Three providers ship today — OpenAI, Anthropic, and OpenRouter — each behind the same generic ProviderAdapter interface, so adding a fourth is one small file, not a rewrite. Credentials are read from server environment variables only; they never reach the browser, get stored in localStorage, or appear in logs, even when a request fails.
The Integrations and Settings pages don't fake this — they show live, server-verified Connected/Not-configured status per provider, so what you see is exactly what's true, not a static mockup pretending to be "connected."
Ten real screens: Overview, Agents, Run detail with a full execution trace, Tasks, Knowledge, Integrations, Analytics, Settings, in-app Docs, and a landing page — all in a distinctive warm, editorial design system (cream backgrounds, teal and terracotta accents) instead of the generic dark AI-dashboard look most agent tools default to.
Fork it, swap in your own domain's demo data, and you have a working foundation for an AI-agent product — not just a UI shell.
AgentDesk — The Control Room for AI Agents
AgentDesk is a control room for AI agents, built to be remixed as a real SaaS starter — not a single-screen chatbo...