Agent map
Tools are described for agents, not guessed from the screen.
Each tool declares its purpose, risk level, confirmation rule, expected output, and recovery path. Visual design helps humans; structured labels guide agents.
WebMCP Challenge Demo
Agent-Native Concierge
A website designed for agents first, then rendered for humans. Users ask for an outcome. The agent brings the right tools, previews likely results, and stops at human approval points.
AI concierge is ready
Voice input is optional. Text input remains the reliable path for every browser.
Human fallback
Manual browsing stays available, but it is not the default.
A user can open the full menu, but the primary flow is concierge: goal first, outcome preview second, tool execution last.
Result-first interface
Choose the result. The agent chooses the tool.
Five-level action control
Agent-safe tools with clear approval boundaries.
Approval sandbox
Dry-run before action
Before any action with outside impact, the agent shows what it plans to do, which tool it will use, why that tool is appropriate, and what the user must confirm.