canvas-chat 0.1.38

A visual, non-linear chat interface where conversations are nodes on an infinite canvas

Canvas Chat

A visual, non-linear chat interface where conversations are nodes on an infinite canvas. Explore topics by branching, merging, and navigating your discussions as a directed acyclic graph (DAG).

Try It Online

No installation required! Try Canvas Chat at ericmjl--canvas-chat-fastapi-app.modal.run.

Bring your own API keys (configured in Settings).

Quick Start

Run Canvas Chat instantly with no installation:

uvx canvas-chat

Your browser will open automatically to the local server.

Features

• Infinite Canvas: Pan, zoom, and navigate your conversations visually
• Branching Conversations: Reply to any message to create a new branch
• Highlight & Branch: Select text within any node to create a highlight excerpt node
• Multi-Select & Merge: Select multiple nodes to combine context
• Context Visualization: See which messages are included in your context
• Auto-Layout: Automatically arrange nodes in a clean left-to-right hierarchy
• Matrix Evaluation: Use /matrix <context> to create cross-product evaluation tables
• LLM Committee: Use /committee <question> to consult multiple LLMs and synthesize answers
• Web Research: Use /research <topic> to generate research reports via Exa
• Web Search: Use /search <query> to search the web via Exa
• Image Analysis: Upload, paste, or drag-and-drop images for multimodal AI analysis
• Markdown & Math Rendering: Full markdown support with LaTeX math rendering (KaTeX) for inline \(...\) and display \[...\] math
• Multiple LLM Providers: Support for OpenAI, Anthropic, Google, Groq, GitHub Models, and local models via Ollama
• Local-First: All data stored in your browser (IndexedDB)
• Export/Import: Save sessions as .canvaschat files

Configuration

Click the ⚙️ Settings button to add your API keys:

• OpenAI: Get from platform.openai.com
• Anthropic: Get from console.anthropic.com
• Google AI: Get from aistudio.google.com
• Groq: Get from console.groq.com
• GitHub Models: Get from github.com/settings/tokens (requires beta access)
• Exa (for search/research): Get from exa.ai

Keys are stored locally in your browser's localStorage.

Admin mode (enterprise)

For enterprise deployments where administrators control API keys server-side:

uvx canvas-chat --admin-mode

This reads model configuration from config.yaml and API keys from environment variables. Users don't need to configure anything - models are pre-configured and credentials are injected server-side.

See Admin Mode Setup for details.

Usage

1. Start chatting: Type a message and press Enter
2. Reply to a node: Click the ↩️ Reply button or click a node then type
3. Branch from text: Select text within a node, then click 🌿 Branch to create a highlight node
4. Multi-select: Cmd/Ctrl+Click multiple nodes to respond to all at once
5. Auto-layout: Click 🔀 to automatically arrange all nodes
6. Search the web: Type /search <query> to search via Exa
7. Research a topic: Type /research <topic> to generate a research report
8. Consult LLM committee: Type /committee <question> to get opinions from multiple LLMs
9. Add images: Upload, paste (Ctrl/Cmd+V), or drag-and-drop images for AI analysis
10. Create a matrix: Select one or more context nodes, type /matrix <context> to create an evaluation table
11. Navigate: Drag the handle (⋮⋮) to move nodes, scroll to zoom, double-click canvas to fit content
12. Export: Click 💾 to save your session as a .canvaschat file

Development

For contributors or local development:

Prerequisites

• Python 3.11+
• Pixi (recommended) or uv

Setup

git clone https://github.com/ericmjl/canvas-chat.git
cd canvas-chat
pixi install

Running

pixi run dev

Open your browser to the URL shown (usually http://127.0.0.1:7865).

Tech Stack

• Backend: FastAPI
• Frontend: HTMX + vanilla JavaScript + CSS
• LLM: LiteLLM (multi-provider support)
• Storage: IndexedDB (browser-local)

Browser compatibility

Canvas Chat works best on Chromium-based browsers (Chrome, Edge, Arc, Brave, etc.). Firefox and Safari have rendering issues with the SVG canvas that prevent full functionality.

License

MIT