Proofboard

Inspiration

Technology decisions are usually split between chat transcripts, spreadsheets, browser tabs, and memory. Research becomes hard to audit, conflicting claims are easy to miss, and an agent can appear more certain than the evidence allows. Proofboard turns the decision page itself into a durable, inspectable collaboration surface.

Why WebMCP

Proofboard is a strong fit for WebMCP because the agent must work with the exact state the person is viewing—not a detached copy hidden inside a chat. Native tools let an agent read human-authored constraints, add structured candidates and source-backed claims, and immediately update the same interface. Human edits are visible to the next tool call.

What people and agents do together

A browser agent can read the room, add candidates, attach validated evidence URLs, and ask Proofboard to surface conflicts and missing fields. It can prepare a shortlist proposal for human review. People can directly edit constraints, inspect every source, exclude or restore candidates, review proposals, and make the final decision.

This division of authority is intentional: Proofboard exposes no finalize_decision agent tool. The Make final decision action exists only in the human interface, and the audit trail distinguishes human and AI changes.

Better user experience

Instead of copying research out of chat, users get one persistent decision room with structured evidence, visible uncertainty, validation errors, provenance, and a human gate. Tool calls update the UI immediately; UI changes are available to subsequent tool reads. The human interface remains fully usable when WebMCP is unavailable.

How it was built

Proofboard is a React/Vinext application deployed on ChatGPT Sites. The UI and all 16 native WebMCP tools share one immutable reducer and browser-persisted state model. After client mount, the app obtains document.modelContext and registers tools with registerTool(...), typed JSON input schemas, read/write annotations, bounded inputs, URL validation, duplicate rejection, and abort-based unregister lifecycle handling.

Read tools cover decision state, constraints, candidates, evidence, conflicts, missing evidence, history, and approval status. Write tools create rooms, constraints, candidates, exclusions, evidence, and shortlist proposals. Applying a shortlist is gated on prior human approval.

Verified behavior

In Chrome's native WebMCP environment, the public site exposed 16 tools and passed a 5/5 native smoke check. A real natural-language tool run added three candidates to the live UI. A later live state contained four source-backed records, one conflict, and nine missing evidence fields, with AI-attributed history and persisted state visible.

The recorded demo does not claim that the pending human-approval step or shortlist application was completed. It shows only the behavior actually captured and verified.

Challenges and learning

The main challenge was preserving one source of truth across direct human edits, asynchronous agent writes, persistence, and strict validation. WebMCP made the product boundary clearer: agents contribute structured work, while consequential approval and final selection remain visibly human-controlled.

Built With

• chatgpt
• react
• sites
• typescript
• vite
• webmcp