ReliefMesh

I was inspired by the amount of good food that can go to waste while community kitchens and shelters still need meals. I wanted to explore how an AI agent could help with this work without making decisions by itself.

ReliefMesh is a food-rescue coordination demo. It shows food offers, partner needs, meal quantities, dietary fit, and pickup windows. An agent can inspect this information and suggest a rescue plan. The plan is shown clearly in the app before approval. A person must review it and give the exact ``confirmation before the app records approval.

I build this app with HTML, CSS, and JavaScript. It uses WebMCP through document.modelContext.registerTool to register five tools: inspecting rescue context, proposing a plan, staging an allocation, approving an allocation, and generating a pickup manifest. I deployed the project on ChatGPT sites and added the headers needed for WebMCP.

The main challenge was making the AI useful without giving it too much control. I separated planning from approval so that the agent can suggest and stage a plan but cannot make a commitment on its own. Another challenge was testing WebMCP in a supported browser. I also checked that document.modelContext was available and that all five tools were registered before recording the demo.

The project successfully registers five WebMCP tools in a supported browser. I am proud of the safety design: the plan is visible before approval, the confirmation must be explicit, and the demo never sends a real-world dispatch or notification.

I learned that a tool should say exactly what it can do and what information it uses. Also, that's how browser permissions, isolation headers, and structured tool inputs affect WebMCP. The biggest lesson was that human approval is not a limitation; it is an important part of a safe workflow.

Built With

• chatgpt-sites
• css3
• html5
• javascript
• webmcp