Inspiration
AI can clarify difficult decisions, but a chat answer is easy to detach from its source, blur facts with assumptions, or accept without a reviewable trail. We wanted a shared workspace where an agent can reason, the application enforces structure and freshness, and the human keeps the final word.
What it does
Thought Amplifier is a subject-independent WebMCP decision workspace for bounded plain-text brain dumps. A compatible browser agent invokes real page-defined tools to read the current source, revision, and fingerprint, then stages a structured amplification: objective, success criteria, facts grounded in exact source excerpts, assumptions, unknowns, options and trade-offs, a recommendation, reversal condition, and next actions.
The result appears in the same visible page. The app validates the payload, binds it to the source revision, records the tool transition, rejects stale writes after the source changes, and exposes no agent approval tool. Only the human-facing UI can approve the exact displayed proposal.
Our demo makes the product test itself: it analyzes the real decision between three challenge ideas—an MCP-readiness auditor, a residential concept coordinator, and Thought Amplifier. The workflow itself is general and supports clarify, decide, plan, diagnose, create, and learn modes.
How we built it
We built a React and TypeScript client with imperative WebMCP registration through document.modelContext.registerTool. Three typed tools share live application state: read_thought_source, stage_amplification, and read_thought_workspace. Closed schemas and runtime validation reject unexpected, malformed, duplicate, or out-of-bounds data. Every staged proposal must match the current source revision and fingerprint; changing the source invalidates the old result. Human approval remains outside the WebMCP tool surface.
The public repository includes architecture and tool-contract documentation, a reproducible real-host verification receipt, a demo runbook, screenshots, tests, and an MIT license.
Challenges
The hardest part was making the integration genuinely useful without turning WebMCP into an unrestricted write channel. We also had to make the manual and agent paths converge on the same state, keep evidence and inference distinct, and ensure a plausible result based on an older source could never silently replace newer work.
Accomplishments
A compatible browser host discovered and invoked all three real page tools in a local end-to-end run. It read the exact live source, visibly staged a complete proposal, read it back, and received STALE_SOURCE when an old payload was replayed after an edit. The registered tool surface contains no approval, publication, submission, or arbitrary-action endpoint.
What we learned
Useful agentic browsing is not just more autonomy. It is precise semantic operations inside software that owns state, validates transitions, exposes provenance, and preserves a human authority boundary. Structured reasoning becomes more credible when the source, transformation, freshness check, and approval are visible together.
What's next
Next: source-span citations, multiple text sources, private document and image ingestion with visible extraction provenance, persistent branches and locks, exports, and permissioned collaboration. This submission deliberately remains a focused prototype: plain text only, session-local state, no embedded model, no autonomous approval, and no claim of production or universal browser certification.
Links
Built With
- chatgpt
- cloudflare
- tailwind
Log in or sign up for Devpost to join the conversation.