Infinite Brainstorm

A desktop infinite-canvas brainstorming app where humans and AI agents work on the same board simultaneously — by editing the same board.json file.

Unlike traditional tools where AI is bolted on through APIs, this is agent-native by design — the JSON file is the API.

That one design choice changes everything downstream: no SDK to integrate, no API keys to provision, no auth dance. Claude Code (or any agent) just reads and writes the same file the canvas is rendering. The desktop app watches for changes and syncs in under 100ms.

Repo

github.com/LucianoLupo/infinite-brainstorm — Rust / Tauri v2 / Leptos.

There's also a bundled Claude Code skill in the repo with the schema docs and layout algorithms, so an agent can be productive on a board immediately without me explaining anything.

Tech stack

• Desktop framework: Tauri v2
• Frontend: Leptos 0.8 (Rust compiled to WASM)
• Rendering: HTML5 Canvas
• File watching: the notify crate
• Link previews: reqwest + scraper for Open Graph metadata

Features

• Infinite canvas with pan / zoom
• Six node types: text, idea, note, image, markdown, link preview
• Directed edges with arrowheads for relationships
• Node metadata: color, tags, status, group, priority
• Real-time external file sync (under 100ms)
• Full undo/redo history (Cmd+Z / Cmd+Shift+Z)
• Cmd+V pastes clipboard images straight to ./assets/
• Obsidian integration — local .md file links render as markdown
• Six board templates: mind map, kanban, flowchart, SWOT, pros/cons, timeline

Why this design

Most "AI-augmented" tools are an editor plus a sidebar where you ask an LLM to do things and then paste results in. The architecture is human edits the doc, agent edits a sidebar.

I wanted the opposite: agent and human both edit the doc. That works only if the doc has a stable, agent-readable format and the renderer reacts to file changes faster than human attention.

board.json + notify + sub-100ms sync is the simplest version of that idea I could ship.

Connection points

• The agent-native pattern is the same one driving agent-orchestrator (CLAUDE.md harness templates as the agent's API surface) and research-orchestrator (shared memory between agents). All three projects are about giving agents a first-class spot at the same workspace as the human, not a sidecar.
• Bundled as the infinite-brainstorm Claude Code skill so any agent in the loop can drive a board.