The problem your project solves

The Problem

Most team knowledge lives in Slack threads, people's heads, and half-finished Notion docs nobody reads. Documentation tools assume you'll sit down and write. You won't.

There are four taxes that kill knowledge sharing: discipline, writing, structure, and maintenance. Every existing tool asks you to pay at least two. Unscatter charges zero.

• New engineers spend weeks reverse-engineering decisions already made
• Critical gotchas live only in one person's head
• Contradictions accumulate silently — two engineers implement the same thing differently
• Nobody knows if what's documented is still true

How you are solving it

The Solution

Unscatter is a local-first, open-source company brain — a macOS menu bar app + web UI that lets anyone dump tacit knowledge (voice / text / file) with zero friction. Claude auto-organizes every dump into structured, queryable notes.

Every dump goes through a single Claude call that deduplicates against existing notes, extracts atomic knowledge units, auto-links related notes, flags contradictions, and updates freshness scores.

Features

• Menu bar app — capture from any app without switching context. Speak for 15 seconds, get a macOS notification: "Created 1 note about pgvector migration"
• Wiki view — auto-organized topic clusters with coverage and freshness badges
• Chat with citations — answers with inline [note_id] references you can click through
• Knowledge graph — force-directed graph with 4 overlays (Coverage, Freshness, Connections, Onboarding Path)
• Telegram bot — capture voice/text from mobile, anywhere
• MCP server — live brain available as tools in Claude Desktop and Claude Code
• Export to Claude Skills — generate SKILL.md files from topic clusters

Architecture

No vector database. No embeddings. No cloud required. The brain lives as plain markdown files + a JSON index. Claude acts as router, organizer, and retriever.

Open source and self-hosted. git clone && docker compose up. Your data never leaves your machine.
The live demo at unscatter-ai.onrender.com runs with seed data — production use is self-hosted.

Use of Genspark

Genspark played a key role in three phases of building Unscatter:

Brainstorming: Used Genspark to rapidly explore the problem space — mapping out why existing knowledge tools fail teams, stress-testing the "four taxes" framing (discipline, writing, structure, maintenance), and exploring architecture tradeoffs between local-first vs. cloud approaches.

Planning: Genspark helped structure the 5-hour hackathon build plan — breaking the product into hour-by-hour milestones with clear gates, and identifying what to cut if running behind schedule.

Logo & branding: The Unscatter logo was generated using Genspark's image capabilities. Fast, high-quality output that matched the dark/purple brand palette perfectly — no design tools needed.

Use of Claude

Claude is the intelligence layer powering every core feature of Unscatter, running on claude-opus-4-7 via the Anthropic Python SDK.

Five Distinct Prompts:

• INGEST: User dumps content → create_note, update_note, add_edge, flag_contradiction operations
• CHAT ROUTER: User asks a question → relevant_note_ids[] (index-only, fast)
• CHAT ANSWER: After routing → answer with inline [note_id] citations
• EDIT: User corrects a fact → full updated note body
• EXPORT: Skill generation → raw SKILL.md content

Key Design Decisions:

AI-as-router: Instead of a vector database, Claude reads the compact index JSON (~10KB) and returns relevant note IDs. The backend loads those notes, and a second Claude call answers. Smarter than keyword search, zero infrastructure.

Structured output discipline: All Claude responses go through parse_claude_json() — a wrapper that strips markdown fences before parsing. Every operation has a typed schema Claude must return.

MCP server: Unscatter ships a FastMCP server so the brain is available as live tools in Claude Desktop and Claude Code — query_brain(), capture(), list_topics().

What is the deployed URL for this project?

https://unscatter-ai.onrender.com