AI diagram generator

AI architecture diagram generator.

Describe the system, or point it at the repository and let it find out for itself.

Download for macOS
v0.1.33 · .dmg · Apple Silicon & Intel
Signed & notarized by Apple · opens without a Gatekeeper warning
sha256 698955a0187bc039f4c74f5d05a9f10fbb27376a45788a0a241d1326b73873c7
Download for Windows instead
$curl -fsSL https://lucidtrain.com/install.sh | sh
E-commerce platform architecture. 13 components across 5 tiers.
E-commerce platform architecture. Produced by the app's layout engine from a semantic graph, not placed by hand.

What an architecture diagram is for

An architecture diagram answers one question: what are the pieces and how do they talk to each other. It is deliberately not a deployment topology, a network diagram or a class model, and trying to make it be all of those at once is the most common reason architecture diagrams become unreadable. The useful version fits in five to fifteen boxes, groups them by tier so the direction of dependency is visible at a glance, and says something about each one rather than only naming it.

How Lucid Train draws one

Describe what you want, or point it at a repository. The agent emits a semantic graph of nodes and edges with no coordinates, and the ELK layout engine computes positions and orthogonal routing. That separation is the reason the output cannot come back with boxes overlapping or arrows crossing through cards, which is the characteristic failure when a language model is asked to place things on a canvas itself. It also means editing is editing the graph: change a component and the layout re-solves rather than drifting out of alignment.

Tiers do the work that layout usually cannot

Grouping components into client, edge, application, data and external bands means the reader knows which way dependency flows before reading a single label. It also makes an anomaly visible: a data-tier component calling out to the application tier is an upward edge and it looks wrong, which is exactly what you want a diagram to do.

From diagram to implementation

One action turns the canvas into a markdown specification listing every component, every relationship and the notes attached to them, then starts a coding turn with it. That is the direction almost nobody else supports: the tools in this category all go from text to picture and stop there.

Local, offline, your models

It runs as a desktop application. Against a local Ollama model the whole thing works offline with no API key and no telemetry, which is the practical difference from every browser-based diagram tool: your architecture never leaves the machine.

What goes in each tier

TierWhat belongs there
ClientBrowsers, mobile apps, CLIs, other systems calling you
EdgeCDN, load balancers, API gateways, anything terminating traffic
ApplicationYour services, workers and the code you deploy
DataDatabases, caches, object stores, queues holding state
ExternalThird-party services you depend on and do not run
InfrastructureObservability and platform components that watch the rest

Try it

shell
$ Diagram the architecture of this repository. Group by tier, show the edges, and note what each component is responsible for.
FAQ

Questions

Can AI generate an architecture diagram from code?

Yes. Point it at a repository and it reads manifests, entry points and the import graph to identify components and the edges between them. The accuracy depends on how much of the architecture is expressed in the code rather than in infrastructure you configured elsewhere.

What formats does it export?

SVG and PNG, plus the Excalidraw scene itself, so diagrams stay editable in any Excalidraw client rather than being locked to this app.

Is there a free architecture diagram generator?

The Lucid Train CLI is free and open source. The desktop app, which is where the Diagram tab lives, is $3/month with a seven-day trial. Model usage is separate: local models cost nothing.

How is this different from asking ChatGPT for a diagram?

A chat model returns Mermaid or an image, and in both cases it is choosing the visual arrangement, which is what produces overlapping boxes and crossed arrows. Here the model only decides what the components and relationships are; a layout engine decides where they go.

Related