Next AI Draw.io

Natural-language diagram generator built on draw.io and your choice of LLM

Freemium· Free hosted demo (rate-limited on the maintainer's key) / self-host free under Apache 2.0 with your own API key (pay upstream to OpenAI, Anthropic, Google, DeepSeek, etc.)Image GenerationModel-agnostic — bring your own (Claude Sonnet 4.5, GPT-5.1, Gemini 3 Pro, DeepSeek V3.2/R1, GLM-4.7, Ollama local, etc.). Public demo runs GLM-4.7 via ByteDance Doubao.

Visit website →

Best for

Engineers and architects who already use draw.io, want a self-hostable natural-language diagram assistant, and are comfortable bringing their own OpenAI / Anthropic / DeepSeek / local Ollama key.

Skip if

Non-technical users who want a polished hosted SaaS with team collaboration, or anyone unwilling to manage API keys, deployments, or occasional manual XML cleanup.

Next AI Draw.io is an open-source Next.js web app that bolts a chat-driven LLM assistant onto the familiar draw.io (diagrams.net) editor. Instead of dragging shapes around, you describe what you want in natural language and the model emits or edits the underlying draw.io XML in place, with the visual canvas staying interactive the entire time. The split-pane UI puts the draw.io editor on one side and a reasoning-visible chat on the other, so you can prompt, tweak, and hand-edit shapes without switching tools.

Beyond text-to-diagram, it accepts image uploads (screenshot an existing diagram or hand-drawn sketch and have the model replicate it) plus PDF and text files as context for architecture, flowchart, sequence, ER, and cloud-topology diagrams. There is explicit support for AWS, GCP, and Azure shape libraries and an 'animated connectors' mode for slide-worthy visuals. History tracking lets you diff and revert AI edits, which matters because LLMs will happily reflow a diagram you liked.

The project is provider-agnostic. It ships adapters for ByteDance Doubao, OpenAI, Anthropic, Google AI, Vertex AI, Azure OpenAI, AWS Bedrock, Ollama, OpenRouter, AIHubMix, DeepSeek, SiliconFlow, ModelScope, SGLang, and the Vercel AI Gateway; the maintainer recommends Claude Sonnet 4.5, GPT-5.1, Gemini 3 Pro, or DeepSeek V3.2/R1 for best XML fidelity. The public demo at next-ai-drawio.jiang.jp runs on GLM-4.7 sponsored by ByteDance Doubao, and lets you paste your own key (stored in the browser, not the server) to bypass shared-quota limits. Deployment targets include EdgeOne Pages one-click, Vercel, Cloudflare Workers, Docker, and plain npm; an admin panel exposes models, access codes, features, observability, and quota without hand-editing .env files. Good pick for engineers who already live in draw.io, want the source, and are willing to bring their own model.

Editor's take

The right shape of tool: it treats draw.io as the source of truth and the LLM as an editor, not the other way round, so you never get trapped in a walled-garden format. Bring a top-tier model — Claude Sonnet 4.5 or GPT-5.1 in our testing — or the connector routing gets ugly fast. Best value is self-hosting on Cloudflare Workers with a DeepSeek or local Ollama key.

— The AI Tool Bible editorial team

Pros

✅ Generates and edits real draw.io XML — the output opens in any diagrams.net editor, no lock-in to a proprietary format
✅ Provider-agnostic: 15+ LLM backends including local Ollama and SGLang, so you can run it entirely offline
✅ Image-to-diagram replication turns screenshots or whiteboard photos into editable draw.io canvases
✅ First-class AWS / GCP / Azure shape support for cloud architecture diagrams
✅ Apache 2.0 licensed with one-click deploy templates for Vercel, Cloudflare Workers, EdgeOne, and Docker
✅ Reasoning panel and edit-history diff make it easy to see what the model changed and roll back
✅ BYO-API-key stored in browser localStorage — no server-side key custody on self-hosted instances
✅ Web-based admin panel manages models, quotas, and access codes without .env editing

Cons

⚠️ Hosted demo is a shared free tier — heavy use requires your own paid model key or self-hosting
⚠️ Quality of the generated XML depends heavily on the chosen model; weaker models mangle layout and connector routing
⚠️ No native collaboration, comments, or team library features that commercial diagramming SaaS offers
⚠️ Small-team open-source project — support is community-only and update cadence tracks a single maintainer
⚠️ Self-hosting still means paying upstream token costs for the good models (Claude, GPT-5.1, Gemini 3 Pro)
⚠️ Complex diagrams often need manual cleanup in draw.io after the AI pass

Use cases

Cloud architecture diagrams (AWS, GCP, Azure)System design sketches from a text specSequence and flowchart generationER diagram drafting from a schema descriptionScreenshot-to-editable-diagram replicationWhiteboard photo digitisationIterative diagram refinement via chatSlide-ready animated connector diagramsSelf-hosted internal diagramming assistant