Comparison

Lucid Train compared with Eraser.

The closest competitor in intent, and the clearest split in approach: hosted and prompt-driven on one side, local and repository-aware on the other.

Download for macOS
v0.1.33 · .dmg · Apple Silicon & Intel
Download for Windows instead·Linux
$curl -fsSL https://lucidtrain.com/install.sh | sh

Lucid Train generates architecture diagrams by reading your actual repository, runs as a desktop application against local models so source code never leaves the machine, and can hand the diagram to a coding agent as an implementation specification. Eraser is a hosted browser tool that generates diagrams from a prompt and offers real-time collaboration. The core difference is that Lucid Train starts from your code rather than from your description of it.

It reads the repository, not your description of it

Eraser's DiagramGPT turns a prompt into a diagram, which means the diagram is only as accurate as your memory of the system. Lucid Train surveys manifests, entry points and the import graph and draws what is actually there, which routinely surfaces components people had forgotten and edges nobody had written down. On an existing codebase that difference is the whole comparison.

Local, offline, your own models

Lucid Train is a native desktop app that runs against any OpenAI-compatible endpoint, including a local Ollama model with no API key and no telemetry. Eraser is a browser product, so anything you paste or connect goes to their servers under their terms. For a large number of organisations that single fact settles which tool is usable.

Diagram to code, and other agents as engines

Eraser goes from text to diagram and stops. Lucid Train converts the canvas into a markdown specification and starts a coding turn that returns reviewable diffs, and it can run that turn on Claude Code, Codex, Cursor or OpenCode as the engine. It is also an MCP server itself, so Claude Code and Cursor can call its diagram tools directly.

Depth in the canvas

Four diagram kinds, six themes, seven tier bands, per-node documentation panels, node states that mark single points of failure, story mode with a narrator, 50 versions of history per diagram, and a vision model that checks the rendered canvas and fixes alignment. That is a canvas built for one job rather than a general drawing surface.

What Eraser is good at

Real-time collaboration, which a desktop application does not attempt. Its diagram-as-code syntax is clean and version-controllable, and being browser-based means a teammate opens a link rather than installing anything. For a team that diagrams together on a call, that matters.

Side by side

Lucid TrainEraser
Generates from a repositoryYes, reads the codeFrom prompts and pasted snippets
Runs fully offlineYes, with local modelsNo
Source code leaves your machineNo, with a local modelYes, hosted
Bring your own modelAny OpenAI-compatible endpointNo
Diagram to code as a specificationYesNo
Runs other agents as enginesClaude Code, Codex, Cursor, OpenCodeNo
Works as an MCP serverYesMCP server available
ExportSVG, 2x PNG, Markdown, .excalidrawSVG, PNG
Version history50 versions per diagramYes
Real-time collaborationNoYes
PriceCLI free and open source; desktop $3/month, 7-day trial, no cardFree tier, then per seat
FAQ

Questions

Is Lucid Train a good Eraser alternative?

For generating and maintaining architecture diagrams from a codebase, yes, and it adds local execution, model choice and the diagram-to-code direction. Eraser keeps the advantage on real-time collaborative editing.

Does Eraser send my code anywhere?

It is a hosted product, so anything you paste or connect goes to their servers under their terms. That is normal for browser tools and is exactly the constraint a local-first desktop app removes.

Can I use Lucid Train from Claude Code or Cursor?

Yes. Lucid Train runs as an stdio MCP server, so its diagram tools are callable from Claude Code, Codex and Cursor directly.

Affiliation

Lucid Train is an independent product and is not affiliated with, endorsed by, or connected to the product described on this page. All names and trademarks belong to their respective owners. Pricing and features were checked on 19 August 2026 and change without notice; check the vendor's own site before deciding.

Related

Last updated