AI that thinks on a canvas

Your whole idea.On one canvas.

Start with a scribble. AI lays out the whole app — every screen, model, and flow — as a map you can see.

A chat gives you a thread to scroll. ThinkCanvas gives you the whole picture — built live with AI, in the messy, hand-drawn way you actually think.

Brainstorm → BlueprintWatch it thinkMCP-native
See the problem first ↓

No credit card · Works with Claude & any MCP agent

The problem

Your best thinking is trapped in a scroll.

AI chat is brilliant — until you need the whole plan back. Then it's 200 messages you have to re-read.

💬 AI chat linear
design a task app with auth
Sure — we'll need a login screen, a task list, and a data model…
what about the onboarding?
Good call. Add a 3-step onboarding after signup…
wait, how does billing connect to auth again?
Billing reads the user id from the session, which comes from…
…and the settings page?
Settings hangs off the dashboard; it also touches the user model…
can you recap the whole thing?
Of course. So far we have: auth (login + session), onboarding…
i've lost the plan 😩
To find the auth flow you scroll, re-read, and lose the thread. The picture only ever lives in your head.
🗺️ ThinkCanvas spatial
The same plan, as a map. Every flow is a node you can see, connect, and jump to — no scrolling.
The solution · live

Watch it design the whole app.

Drag the finished map to explore it — or hit replay and watch the AI build the whole blueprint, node by node.

AI-note-takingMake your own
ingestreadyon demandsplit intopassagesvectorizeeach chunkindex forsearchquery-timeside-by-sidefeedscontextpowerszoom into "AIAssistant"zoom into reading UXassistant lives in theright panelREST /WebSocketorchestratesauth + dataparseuploadsvectorizestoreprompt with contextRAGretrievalzoom into clientzoom into backend
AI Note-Taking App — Core Flow
flow →
Upload PDF
AI understands the doc
Read with AI at your side

AI Assistant

AI Assistant Functions
mindmap

Ask the AI

Summarize sections or the whole doc

Explain complex passages in plain language

Q&A — ask specific questions about the content

Cross-reference across pages / connect ideas

Extract key takeaways, definitions, or action items

How AI Understands a PDF
flow ↓
PDF file

Chunking engine

Embedding model

Vector store

RAG retrieval

Chunking strategy is everything — too small = lost context, too large = irrelevant retrieval. Overlapping chunks +…
Split-Panel Reading Experience
flow →
PDF reader panel

Highlight text → ask about just that passage

Context-aware: AI sees what you're reading

AI chat panel
AI Note-Taking App — Architecture & Experience
flow ↓
Tech Stack — Three Tiers
flow ↓
iPad Client
API Layer
AI Pipeline
iPad Client — The Fork
tree ↓

Two viable paths

Next.js + Tailwind — responsive web, works in Safari & as PWA (add to Home Screen)

SwiftUI native — best PDF perf, Apple Pencil, Slide Over multitasking

React Native — shared codebase with web but PDF rendering is painful

Start PWA (fastest to ship) → native when PDF perf or Apple Pencil becomes the bottleneck

Backend Stack
flow →

FastAPI (Python)

Supabase

PyMuPDF / Unstructured.io

OpenAI text-embedding-3-small

pgvector (in Supabase)

Claude / GPT-4o for Q&A

iPad = offline matters. If the user reads on a plane or commute, the PWA path needs a local embedding model (e.g.…
Tech Stack for AI Note-Taking App (iPad-first)
flow ↓
📝Second Brain — Business Plan (July 2026).md24 KB

Replay it to watch the thinking happen. On your own board, this is your idea becoming your app.

Why a canvas

Scan, don't read.

Reading is linear. Seeing is instant. A canvas hands you the whole system at a glance — so you spot what matters and move.

See the whole system at once

Every screen, model, and dependency on one surface — not paged out across a transcript.

Jump anywhere, no scrolling

Click a node to go straight to it. Your plan is a place you navigate, not a wall you re-read.

Spatial memory beats a log

You remember where things are. A map sticks; message #147 does not.

Agent-native

Built for your agents.

ThinkCanvas speaks MCP — your agent designs on the same board you do.

Point Claude at your canvas. It reads and writes it directly.

No copy-paste, no screenshots. Connect ThinkCanvas over MCP and your agent adds nodes, draws edges, and maps out your app on the real board — then you take over on the same surface.

Claude-nativeMCP read + writeYour board = your agent’s board
Pricing

Start free. Upgrade when it clicks.

MonthlyYearly save ~17%
Free
$0/mo
For trying the canvas.
For trying the canvas and small ideas.
  • Up to 3 boards
  • AI brainstorm → graph
  • Watch it think replays
  • All layout types
Get started
Most popular
Starter
$12.99/mo
Billed monthly · 14-day trial
For builders designing real apps.
  • Unlimited boards
  • Bring your own agent (MCP)
  • Priority AI + doc / image generation
  • Share links + replays
Start 14-day trial
Team
Let's talk
On the roadmap.
For teams designing together.
  • Everything in Starter
  • Shared team boards
  • Real-time collaboration
  • SSO + admin
Talk to us