AgentDesk

AgentDesk is a front desk for AI agents that any website adds with one line: app.use(agentdesk()), plus a MongoDB URI and an LLM key. It starts knowing nothing. Agents find it on their own (llms.txt, a Link header, /.well-known/mcp) and talk to it over MCP or plain HTTP with five fixed tools: about, ask, list_actions, do, done. Every visit teaches it. It taps the requests an agent makes on the site, and when the visit ends a reflection loop turns what happened into facts, reusable actions (replayable request templates) and proposed guardrails. When it could not help, it reads its own site (llms.txt, pages, forms) and learns from that. The next agent gets the answer, or finishes the task in one call instead of clicking through the site. Autopilot is on by default: learned actions go live right away, anything that spends money waits for approval, every unreviewed run is audited, and a circuit breaker pauses an action after two failures. The developer panel shows each visit, what was learned, the learning curve and the audit log, with one-click approve or block. Demo: Harbor Goods, a small store with AgentDesk installed. Agent 1 hears "I don't know yet"; about 20 seconds later agent 2 adds to cart in one call. Live: https://agentdesk-mv4ds7c7j01-prod.cloud.modelence.app