Infinite Canvas: A Local-First AI Image Workbench That Skips the Server

What Infinite Canvas Actually Solves

Infinite Canvas targets a specific pain: generating images with AI is usually a one-shot affair inside a chat window. You type a prompt, get an image, and then you start over if you want to iterate. This project instead puts image generation inside an infinite canvas where you can arrange nodes, connect them, and refine results in place. The intended user is someone who explores visual ideas continuously, not someone who just wants a single picture. The README describes it as a workbench for image creation that combines canvas orchestration, AI image generation, reference image editing, a conversation assistant, a prompt library, and asset storage in one interface. It is for people who want to keep context while they iterate, which is a real gap in most chat-based image tools.

How the Data Flow Works: Browser First, No Backend

The most distinctive architectural choice is that the browser front end talks directly to your OpenAI-compatible API. There is no intermediary server for API calls. The README states that AI API keys, Base URLs, canvases, assets, and generation records are stored locally in the browser by default. This means your prompts and images do not leave your machine except to reach the API endpoint you configure. The canvas itself supports multiple projects, node dragging and zooming, connections, a minimap, undo and redo, and import/export. The assistant feature talks to the selected node and its upstream nodes, then inserts results back into the canvas. That is a concrete loop: you select a node, ask a question or generate, and the output becomes a new node. The local Agent goes further: it connects to Codex or Claude Code via MCP, letting an external agent operate on the canvas. This is a browser-based tool that treats the canvas as a stateful workspace, not a static image viewer.

Getting It Running: Commands and Configuration

Setup is straightforward if you have the toolchain. For local development, the README gives these commands: git clone git@github.com:basketikun/infinite-canvas.git, then cd infinite-canvas, cd web, bun install, and bun run dev. That assumes you have Bun installed, which is a specific dependency. For a more portable deployment, you can use Docker: git clone the same repository, cd infinite-canvas, then run docker compose up -d. The default port is 3000, so you access it at http://localhost:3000. After first launch, you go to the configuration in the upper right corner and fill in your OpenAI-compatible Base URL and API Key. If your API does not match the default OpenAI calling convention, the README says you can customize the generation and video call scripts. That is a practical escape hatch for the many API providers that are not perfectly OpenAI-compatible. The project also offers a Render deployment button, which is visible in the README links, so you can host it without a local machine.

The Plugin System and Custom API Calls

Infinite Canvas includes a plugin system that lets you install, enable, update, and uninstall remote node plugins via URL. There is a TypeScript SDK for developing your own canvas node plugins. This is a significant extension point: you are not stuck with the built-in node types. The README also mentions custom API invocation: you can define how image and video generation calls are made, which is meant to adapt to various relay stations and self-hosted services. That flexibility is important because the project's sponsor list includes API relay services like Infistar.ai that offer access to many models, but the core requirement remains OpenAI compatibility. If your provider uses a different protocol, you will need to write a custom script. The documentation does not give an example of such a script, so you should expect to read the source code or experiment on your own.

A Real Limitation: Pre-Release Instability

The README carries a caution notice that the project is in the development stage and does not guarantee historical data compatibility. It explicitly says that various local storage formats may be adjusted directly. That is a serious constraint for anyone who wants to build a lasting workflow. Your canvases, prompts, and generation records live in IndexedDB, and a future update could break them. The README advises that if you need a stable maintained branch, you should fork the project and develop independently. That is not a typical caveat; it is a warning that the project may change under you. For a tool whose whole value is accumulating a visual workspace, losing that workspace is a real risk. If you are evaluating this for professional use, this alone might disqualify it until the project stabilizes.

The License Contradiction and What It Means

The repository metadata states the license is AGPL-3.0, but the README's open-source license section says the project uses the MIT License and allows anyone to use, copy, modify, distribute, relicense, and use it commercially, including in closed-source products. That is a direct contradiction. The metadata at the top of the repository page, which is what GitHub and package managers read, says AGPL-3.0. The README text says MIT. You cannot trust both. If you plan to use this project in a commercial product, you must resolve this discrepancy before relying on it. The AGPL would impose copyleft obligations on network services, while MIT would not. This is not a legal opinion, but it is a factual inconsistency that should make any adopter pause. Check the LICENSE file in the repository and any recent commit that may have changed the license, because the README may be outdated.

Alternatives: ComfyUI and Dedicated Node Tools

The most direct alternative is ComfyUI, a node-based AI image generation tool that runs locally and offers fine-grained control over generation pipelines. The difference in approach is that ComfyUI is a desktop application that manages its own workflows and model files, whereas Infinite Canvas is a browser-based tool that relies on external APIs. ComfyUI gives you offline generation and deterministic workflows, but it requires GPU setup and model management. Infinite Canvas is lighter to start because you only need a browser and an API key, but you depend on a third-party API for every generation. Another alternative is using a plain chat interface like ChatGPT or a dedicated image generation web app, but those lack the canvas and node-based iteration. If you need offline control, ComfyUI is the better fit. If you want a lightweight, API-driven canvas with minimal setup, Infinite Canvas has a niche, but the instability and license confusion are real costs.

Maintenance and Upgrade Cost: What to Expect

The project has recent releases, with v0.16.0 pushed on 2026-08-18, so it is actively maintained. The version numbers are still below 1.0, which matches the pre-release warning. Upgrade cost is potentially high because of the data compatibility risk. Each release could change the IndexedDB schema or the canvas file format, and you may need to migrate or lose data. The README does not describe any migration path. The plugin system also means that third-party plugins may break with updates. If you fork the project, you take on the maintenance burden yourself. The project's homepage is canvas.best, and the repository is not archived, so development is ongoing. But ongoing development in a pre-1.0 project means frequent changes. Budget time for testing each upgrade and backing up your browser's IndexedDB before pulling a new version.

Editorial conclusion

Adopt Infinite Canvas if you are a developer or power user who wants a self-hosted, browser-based tool for AI image generation and iterative visual exploration, and you are comfortable with the project's pre-release status. Do not adopt it if you need stable data formats, long-term maintenance guarantees, or a production-grade workflow. Before adopting, verify that your AI API is OpenAI-compatible or that you can write custom call scripts, and check the repository's recent commits and release notes for breaking changes. Fork the project if you plan to rely on it, as the README explicitly warns that local storage formats may change without notice.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes