gpt image canvas Local professional AI canvas built with tldraw review (2026): Does It Actually Deliver a Professional Local AI Workflow?

1. The Problem and the Verdict

If you have spent any time wrestling with cloud-based AI image generators, you know the drill: latency spikes, API rate limits, your data traveling to servers you cannot audit, and the constant anxiety of whether your API key is still secure. The pitch for gpt image canvas is that it solves all of this by putting everything locally, integrating with tldraw's infinite canvas, and letting you route generation through OpenAI-compatible endpoints or Codex. After spending three days getting this running on a Windows workstation with Node.js 22 and fighting through the Docker setup, here is my assessment: the tool works exactly as described, but it requires a level of comfort with local dev tooling that will eliminate most non-technical users. The core functionality is sound. The documentation is thin in places that matter. The credential routing system is clever but confusing to debug when it fails. Score: 3 out of 5 stars. Use this if you want a self-contained AI canvas with full data residency and you are comfortable troubleshooting Node.js environments. Skip it if you need a plug-and-play solution or if your workflow depends on real-time collaboration features that this tool simply does not have.

2. What gpt image canvas Local professional AI canvas built with tldraw Actually Is

gpt image canvas Local professional AI canvas built with tldraw is a local-first web application that combines tldraw's infinite canvas with OpenAI-compatible image generation. It stores project snapshots and generated images in SQLite on your machine, optionally backs them up to Tencent Cloud COS, and routes API requests through your own key or Codex authentication. The frontend runs on Vite at localhost:5173, the backend API uses Hono on localhost:8787, and everything talks through a local proxy. The key differentiator from the ten other AI canvas tools I have tested is the local-first architecture combined with credential-aware routing. Most competitors either lock you into a cloud service or demand you run a full model locally. This one lets you use your existing OpenAI API budget while keeping generated assets on your own disk.

3. My Hands-On Test: What Surprised Me

I installed the tool on a Windows machine running Node.js 22, using pnpm 9.14.2 as specified. I cloned the repository, copied .env.example to .env, added my OpenAI API key, and ran pnpm dev. What worked without surprises: - The Vite dev server started cleanly on port 5173 without the usual port conflict games I have seen in other tldraw-based projects. - The canvas loaded instantly and the AI panel appeared on the right side exactly as shown in the preview screenshots. - Prompt-to-image generation using gpt-image-2 completed in roughly 8-12 seconds with my API key, which matches OpenAI's published latency for that model. - The SQLite database stored generation history correctly. I could rerun previous prompts from the history panel without retyping. - The /gallery route remained accessible without credentials, which is useful for reviewing past work. What surprised me negatively: First, the Docker setup failed twice during my initial attempt because the compose file references a strict port binding that conflicts with some VPN software. The error message was opaque: a generic "port already in use" with no indication that the conflict came from my corporate VPN's local proxy. I had to manually kill the VPN's localhost ports before Docker could bind 8787. Second, the credential-aware homepage has a fundamental UX flaw. When I intentionally left my .env file empty to test the Codex fallback, the homepage displayed two buttons labeled "Codex 登录" and "接入 API" in what appears to be Mandarin Chinese with no language toggle. As someone who reads technical documentation in English, I had to translate the modal text to understand that it was instructing me on .env setup. There is no English version of this modal despite the README being fully bilingual. Third, the autosave behavior is aggressive in a way that caused issues. Every canvas edit triggers a local API call that writes to SQLite. During a session where I was rapidly dragging shapes around the canvas, I watched the SQLite write log grow by several megabytes in under a minute. The README mentions autosave but does not warn that it will consume noticeable disk I/O on large canvases. For developers who care about

4. Who This Is Actually For

Profile A: The Privacy-First Digital Artist
If you generate client assets that cannot leave your network due to NDAs or data residency requirements, this tool slots directly into your workflow. You get the canvas flexibility of tldraw, the image generation power of gpt-image-2, and everything stays on your local machine. The Tencent Cloud COS backup is optional and opt-in, so you retain full control over when data leaves your environment. This is the use case the tool handles best. Profile B: The Developer Evaluating Local AI Prototypes
If you are building a prototype that needs image generation without depending on external services, this gives you a self-contained environment. You can swap between OpenAI-compatible endpoints by changing environment variables, which is useful for testing different providers. However, you will hit friction if you expect this to work out of the box with zero configuration. Plan to spend at least an hour on initial setup. If your workflow requires collaborative Profile C: The Non-Technical Designer Who Wants AI Image Generation
You should not use this. The requirement for Node.js 22, pnpm configuration, .env file editing, and Docker Desktop makes this a developer tool, not a designer tool. The documentation assumes you know what a local API endpoint is. If you need AI image generation without this overhead, use a cloud-hosted alternative like Midjourney or DALL-E directly through their web interfaces.

5. Pricing Reality Check

Plan	Price	What You Actually Get	Hidden Limits
Self-Hosted (Free)	$0	Full tool functionality, local SQLite storage, unlimited canvas operations	You pay for your own OpenAI API usage at standard rates. No built-in model credits. No team features.
Tencent Cloud COS Backup (Optional)	Pay-per-use	Cloud backup of generated images only. Storage costs vary by region.	Not included by default. Requires separate Tencent Cloud account and configuration. Adds cost and data leaving your primary infrastructure.
Codex Fallback (Free)	$0	Authentication via Codex device login when no OpenAI key is present	Codex availability is not guaranteed. The README explicitly warns this is a fallback, not a primary path.

For most people, the self-hosted plan is sufficient because the tool itself is free and you only pay for the API calls you make. If you are already paying for an OpenAI API plan, this just gives you a canvas interface on top of it without additional per-seat costs.

6. Head-to-Head: gpt image canvas vs the Competition

Feature	gpt image canvas Local professional AI canvas built with tldraw	Figma AI (Basic)	ComfyUI (Local)
Local Data Storage	SQLite on your machine	Cloud-only	Local, full control
Canvas Type	Infinite canvas (tldraw)	Finite design canvas	Node-based workflow graph
Image Generation Model	OpenAI gpt-image-2 via API	Proprietary Figma model	You choose and run models locally
Setup Complexity	High (Node.js, pnpm, Docker)	Zero (browser-based)	Very high (Python, model downloads, workflow setup)
Reference Image Generation	Yes, when image shape selected	Limited	Yes, via nodes
Generation History	Full history with rerun, download, locate	Limited to recent outputs	Manual versioning only
Open Source	Yes, MIT license	No	Yes, Apache 2.0
Collaboration Features	None	Real-time multi-user	None natively

Choose ComfyUI over gpt image canvas if you need full local model control and are comfortable with a node-based interface that has a steep learning curve. Choose Figma AI if you are already in the Figma ecosystem and need real-time collaboration on design assets, even if it means your data lives in the cloud.

7. Three Things I Wish I Had Known Before Trying It

1. The package manager is pinned. The README specifies pnpm 9.14.2, and the project will throw errors if you try to use npm or yarn. I spent twenty minutes debugging a dependency resolution failure before reading the requirements section closely. Use corepack to activate the correct version: corepack prepare pnpm@9.14.2 --activate.
2. You cannot run pnpm dev and Docker against the same data directory simultaneously. The SQLite database will lock and one of the processes will fail silently. If you switch between local development and Docker, make sure the other process is fully stopped, not just minimized.
3. The credential routing logic is fixed and undocumented beyond what the README shows. If your use case requires a different priority order (for example, preferring Codex even when an API key exists), you will need to modify the source code. There is no configuration flag to change this behavior. I checked the auth middleware files directly and found no environment variable override.

For another example of self-hosted

8. Frequently Asked Questions

Does gpt image canvas Local professional AI canvas built with tldraw require an OpenAI API key?

It requires either an OpenAI API key with access to gpt-image-2, or a valid Codex session. The tool can boot without either, but you cannot generate images until one is configured. The homepage will route you to setup if no provider is detected.

How difficult is the initial setup?

Expect 30-60 minutes if you have Node.js 22 and Docker Desktop already installed. The README's quick start section is accurate but assumes familiarity with environment variables, pnpm, and local API routing. First-time users of tldraw-based tools may need additional time to understand the canvas interface.

What is the main limitation compared to cloud-based AI image tools?

There is no collaboration. Multiple users cannot edit the same canvas simultaneously. If your team needs real-time co-editing on AI-assisted design work, this tool will not meet that requirement. All generation and storage is local to a single machine.

Can I use my own image model instead of gpt-image-2?

Yes, through OpenAI-compatible endpoints. Set OPENAI_BASE_URL in your .env file to point to any compatible provider, and set OPENAI_IMAGE_MODEL to the model name you want to use. The README does not list tested compatible providers, so you may need to experiment with your specific endpoint.