Engines

Two ways to run on a subscription you already pay for.

Neither route asks you for an API key. One of them is a supported integration and one of them is interop; the difference matters.

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

There are two routes and they have different risk profiles. The first is to select Claude Code or Codex as the engine, which spawns that vendor's own CLI as a local process using the login it already has; nothing is proxied and Lucid Train never sees a credential. The second is to sign the bundled engine in with Claude or with ChatGPT directly, importing an existing Claude Code or Codex CLI login or running the OAuth flow in your browser, after which the bundled engine runs on that subscription instead of a pay-per-token key. The second route talks to undocumented endpoints and is reverse-engineered interop subject to the vendors' terms.

Route one: drive the vendor's own CLI

This is the conservative option and the one to prefer if you have a choice. Selecting Claude Code, Codex, Cursor CLI or OpenCode as the engine spawns that program from your PATH and speaks its protocol. You are running the vendor's client, doing what the vendor built it to do, and Lucid Train is the thing rendering its output and gating its edits. Nothing about your relationship with that vendor changes.

What that means for the credential

On this route Lucid Train never sees the credential. It looks for the binary on PATH, spawns it as a child process, and speaks its protocol. Authentication is whatever that CLI already did when you signed into it, so the entitlement stays between you and the vendor and nothing is proxied, replayed or re-signed.

Route two: sign the bundled engine in

There is a second route, and it is worth being straight about what it is. The bundled engine can sign in with Claude or with ChatGPT directly, importing an existing Claude Code or Codex CLI login or running the OAuth flow itself, and then run on that subscription rather than on a pay-per-token key. This talks to endpoints the vendors do not document and that expect their own client identifiers. It is reverse-engineered interop, subject to Anthropic's and OpenAI's terms, and it is not officially supported or endorsed by either of them. It works today and could stop working whenever they change something.

Why route two exists at all

The bundled engine is the only one that can route different roles to different models, use the diagram tooling natively and connect a new Grafana instance. Running it on a key you pay per token for is a real cost when you already pay a subscription every month. That is the whole motivation, and it is why the option is there despite the caveat above rather than because the caveat does not apply.

It shares the token file on purpose

The ChatGPT route writes refreshed tokens back to the same auth.json the Codex CLI uses. Refresh tokens there are single-use, so two clients keeping separate copies would invalidate each other on the next refresh and you would be signed out of one at random. Sharing the file keeps both working, which is a small detail that prevents a confusing failure.

The option with no caveat at all

The bundled engine also runs against local Ollama models with no key, no account and no network, and against OpenRouter or any OpenAI-compatible endpoint with a key you own. If the terms question above matters to you or to your employer, those are the routes that raise no question, and local models raise none about data either.

The routes compared

	Drive the vendor CLI	Sign the bundled engine in
Needs an API key	No	No
Lucid Train sees the credential	No	Yes, stored locally
Uses documented interfaces	Yes	No
Officially supported by the vendor	It is their own client	No
Model routing across roles	No	Yes
Can connect a new Grafana	No	Yes
Could break without warning	Unlikely	Yes

FAQ

Questions

Is subscription sign-in officially supported by Anthropic or OpenAI?

No. It works against endpoints neither company documents for third-party clients, and both auth modules in the source say plainly that it is reverse-engineered interop subject to those terms. Treat it as such.

Which route should I use?

If you can, drive the vendor's own CLI as the engine: you are running their client and no interop question arises. Sign the bundled engine in only when you specifically need what the bundled engine does, such as routing roles to different models.

Is there an option with no terms question at all?

Yes, two. Local Ollama models need no key, no account and no network. Your own API key against OpenRouter or any OpenAI-compatible endpoint is a normal documented use of that key.

Where are the tokens stored?

Locally. The Claude route writes to a file in your Lucid Train config directory with owner-only permissions; the ChatGPT route writes back to the Codex CLI's own auth.json, deliberately, because its refresh tokens are single-use and two copies would invalidate each other.

Run Lucid Train on Claude Code
The same binary you already run in a terminal, driven from a window that also has your diagrams, your dependency audit and your dashboards in it.
Run Lucid Train on the Codex CLI
JSON-RPC to a local app-server, and the only engine besides Claude Code that can also drive the design canvas.
Run Lucid Train on the Cursor CLI
Cursor's headless agent, with the diagram canvas and the approval gate wrapped around it.
Run Lucid Train on OpenCode
The open-source option, and the only engine that talks HTTP rather than stdio.

Last updated 22 August 2026