Use case

Diagramming without an internet connection.

Every other tool in this category is a web application. If your code cannot go to a browser, that rules all of them out.

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

The situation

Defence, healthcare, finance and any organisation with a serious data classification policy face the same constraint: source code and architecture cannot leave the network, and often the machine has no route out at all. Every hosted diagram tool is unusable under that constraint, not because of a policy disagreement but because the network connection does not exist.

What fully offline means here

A desktop application running against a local Ollama model. No API key, no telemetry, no outbound connection required for anything: the models are local, the icons are bundled rather than fetched, and the diagrams are files on your disk. Credentials, when you use any, live in the OS keychain. Disconnect the machine and nothing changes.

Icons bundled, not fetched

A small detail that decides whether this actually works: service logos ship with the application. A tool that fetches icon assets renders a diagram full of blank boxes on an isolated network, which is the kind of failure that only appears after deployment.

The quality tradeoff, stated honestly

Local open models are behind frontier hosted ones, and for diagram generation the gap is narrower than for open-ended work, because the task is structured: read files, emit a graph. It is a real gap nonetheless. If your constraint permits a hosted model, you will get somewhat better results; if it does not, this works and the alternatives do not run at all.

Fully local

shell
$ ollama pull qwen3-coder
# Point Lucid Train at the local endpoint. No API key, no network.
FAQ

Questions

Does it work with no network at all?

Yes. Local models, bundled icons, local files. There is no call it needs to make.

Is there any telemetry?

None unless you opt in. The desktop app collects nothing by default and that is not configurable-by-accident behaviour, it is the default.

Which local models work?

Anything Ollama serves, and any OpenAI-compatible endpoint including a self-hosted inference server. Coding-oriented models do noticeably better at reading a repository.

How much worse are local models here?

Noticeably behind frontier models on subtle reasoning, close enough on structured extraction that a generated diagram is still a useful first draft. Given that the alternative under this constraint is drawing by hand, the comparison is favourable.

Related