InquisitiveD

Talk or type — Inquisitive's agent plans over 9 WebMCP tools to sketch architecture diagrams live on Excalidraw, while any WebMCP agent can inspect, edit, and extend the same canvas.

Comment

Inspiration

Describing complex architecture in text is simple, but visually mapping several projects over a period of time, it remains a major bottleneck. While typical AI agents struggle to parse complex web UIs, WebMCP changes the paradigm by enabling web apps to expose explicit, structured tools directly to models. I built Inquisitive to use for my daily project review, design, sharing, and even technical discussion

What it does

Inquisitive is a technical diagramming app. Through text or voice commands like "diagram an event-driven payment flow," an orchestrator plans across 9 WebMCP tools registered on navigator.modelContext to construct diagrams live on an Excalidraw canvas. The agent inspects canvas topology, modifies specific nodes, appends shapes, captures visual snapshots, and offloads structural generation to a subagent. Meanwhile, users can sketch, speak via Gemini Live audio, lock views, and export multi-format session reports. Any external WebMCP client gains identical programmatic control over the canvas.

How we built it

The stack runs on React 19, TypeScript, Vite, and @excalidraw/excalidraw. Nine strictly-typed WebMCP tools bridge to the scene via an ActiveCanvasBridge. A Gemini function-calling orchestrator (gemini-2.5-flash-lite, $\le 5$ tool turns) delegates vector generation to a subagent bound by a rigorous JSON schema, pairing a lightweight catalog (73+ stencils in ~550 tokens) with client-side vector hydration. Ollama (gemma4:31b-cloud) mirrors the identical tool loop over a Cloudflare Pages /api/proxy. Voice interaction taps Gemini Live audio with a Web Speech API fallback, complemented by an AES-GCM-256 WebCrypto vault, IndexedDB history with Supabase RLS sync, and resilient JSON repair.

Challenges we ran into

Testing webmcp, fixing duplicate MCP tool names, adding an audio-to-audio pipeline and later removing it for unsatisfactory performance; one-way audio is present

Accomplishments that we're proud of

Just spent 2 and a half days visualizing and finishing up to a minimal satisfaction level; for me, the most important thing is that I should be able to use it just after finishing it : )

What we learned

What's next for Inquisitive

Not sure, but it will be an extensive part of my daily routine

Built With

Share this project:

Updates

Submission history