Dal testo al diagramma. Disegnato dall'IA, sul tuo dispositivo.

Diagram Studio is a free AI flowchart and mindmap generator that runs entirely in your browser. Describe what you want — a signup flow, a database schema, a sprint workflow, a release branching strategy — and a small on-device language model writes the Mermaid code, renders it live as it streams, and lets you refine the result in a code editor before exporting as SVG or PNG.

It is built for the situations where cloud diagram tools are the wrong answer: drafting a quick sequence diagram from a meeting note, generating a private ER schema for a client system you cannot put on someone else's server, or producing a publishable flowchart without signing up for anything. Nothing leaves your device.

How it works

Two engines sit behind the same interface. The first is Chrome's built-in AI (the LanguageModel API, available in Chrome 127+ and other Chromium browsers that ship Gemini Nano). It runs the model the browser already has embedded, so there is nothing to download — generation starts immediately and stays on-device. The second engine is a local Qwen2.5 1.5B model loaded through Transformers.js on WebGPU (with a WebAssembly fallback when WebGPU is unavailable). It is a one-time download of roughly 1 GB, cached by the browser for offline use afterwards.

Whichever engine runs, the output is a Mermaid code block. A small regex-based auto-fixer repairs the common slips a 1.5B model makes — unbalanced subgraph ends, missing diagram-type keywords, unclosed quotes in node labels — before the code reaches the renderer. Mermaid then parses and renders the code, with a streaming-tolerant mode that silently ignores mid-generation syntax errors and a final render that surfaces real errors with a clear message.

Diagram types and themes

Eight diagram types are supported: flowchart, mindmap, sequence, class, state, entity-relationship, Gantt and git graph. You can pin a type explicitly, or leave the selector on Auto and let the model pick the most appropriate representation from your prompt — a request about "database tables" tends to land on an ER diagram, a request about "user journey" on a flowchart, a request about "message exchange" on a sequence diagram.

Four built-in Mermaid themes (Default, Neutral, Dark, Forest) recolour the rendered diagram instantly without re-running the model. The code itself is theme-agnostic, so you can switch themes mid-edit without losing your work.

Editing and export

The Mermaid code is always exposed in an editor next to the canvas. Every keystroke re-renders the diagram live. You can rename nodes, rewire edges, add subgraphs, swap the diagram type, fix a label the model got wrong — the editor is the source of truth, not a read-only artifact.

Export supports SVG (vector, infinite resolution, ideal for web and print) and PNG at 2× or 4× raster resolution. Both are generated client-side from the rendered SVG via canvas rasterization — no server round-trip, no watermark, no account. SVG preserves interactivity and scales cleanly into a documentation site or a slide deck; PNG is the right choice when you need a single self-contained image for chat, email or a slide.

Limits & requirements

The local Qwen model needs a reasonably modern device for acceptable speed — WebGPU support matters, and 1.5B parameters still means a multi-second generation for a complex flowchart. The Chrome built-in AI path is faster when available but only exists on Chromium browsers that ship Gemini Nano. If neither is available, the tool still works as a pure Mermaid editor: write the code by hand, get live rendering and the same export options.

A 1.5B model is not a frontier model. It will occasionally produce a syntactically wrong construct that the auto-fixer cannot repair, or mislabel a node. The live editor exists precisely for this case — treat the AI output as a first draft, not a final artifact. For diagrams with strict semantic requirements (UML compliance, BPMN naming), expect to do some manual cleanup.

Privacy

Your prompt and the generated Mermaid code never leave your browser. There is no server-side generation, no logging, no analytics on the prompt content, and no account. The only network request the tool ever makes is the one-time model download from a CDN — and once the model is cached, even that goes away. The built-in AI path makes no network requests at all.