INKFORGE

Visual Thinking with AI — a bidirectional visual dialogue loop between humans and LLMs.

You sketch, the AI responds visually. You iterate, it refines. No more text-only conversations for visual ideas.

     ┌──────────┐                      ┌──────────┐
     │  HUMAN   │ ────── sketch ─────► │    AI    │
     │          │ ◄───── sketch ────── │          │
     └──────────┘                      └──────────┘
           │                                │
           └──────────── ITERATE ───────────┘

INKFORGE Lite

A lightweight, standalone visual canvas with an HTTP API. Any AI agent can see what you draw and respond with SVG — no vendor lock-in.

• Draw with mouse or stylus (freehand, erase, color picker)
• HTTP API at localhost:8765 — simple REST endpoints
• MCP Server — integrates directly into Claude Code as a tool
• Cross-platform — runs on Windows, macOS, Linux (Python + HTML5)

Quick Start

cd inkforge-lite
pip install -r requirements.txt
python inkforge.py

The canvas opens. An API is available at http://127.0.0.1:8765.

With Claude Code (MCP)

cp .mcp.json.example .mcp.json
# Edit .mcp.json to point to your Python path if needed

Claude Code can then use inkforge_open, inkforge_snapshot, inkforge_draw, and other tools to interact with the canvas.

API

Endpoint	Method	Description
/status	GET	Health check
/snapshot	GET	Canvas as base64 PNG
/changed	GET	Check for user modifications
/draw	POST	Inject SVG onto canvas
/clear	POST	Reset canvas
/ack	POST	Acknowledge changes

See inkforge-lite/README.md for detailed API docs and examples.

What's Next

The current canvas is a flat surface. The next step is to separate the visual loop into three independent layers:

  ┌─────────────────────────────────────────────┐
  │  FOND       what you're working on          │
  │             UML, wireframe, todolist, math   │
  ├─────────────────────────────────────────────┤
  │  FORME      how it's represented            │
  │             SVG, Mermaid, ASCII, code        │
  ├─────────────────────────────────────────────┤
  │  INTENTIONS what you want to do             │
  │             add, modify, delete, annotate    │
  └─────────────────────────────────────────────┘

This separation lets the AI understand what you're sketching, how to represent it, and what action you're requesting — instead of treating every stroke as raw pixels.

Vector canvas — Replace the raster canvas with SVG strokes (via perfect-freehand). A simplification slider lets you go from raw freehand to clean, simplified polylines. The AI receives vector data instead of pixels — lighter, more semantic, easier to reason about.

  raw strokes ──────────────────────── clean vectors
  ●∿∿∿∿∿●        simplify ──►         ●────────●
  (raster PNG)                         (SVG paths)

Autonomous loop — Today the user must tell the AI "I drew something". The goal is a fully reactive loop: the canvas detects when you stop drawing and triggers the AI automatically — no text prompt needed, pure visual dialogue.

See INKFORGE.md for the full project vision.

License

MIT