Coding agent

Run a coding agent on local models.

For when the constraint is not cost but that the code cannot leave the building.

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

What running locally actually requires

Three things, and most tools fail at least one. The application has to be a desktop binary rather than a web app. It has to accept an arbitrary model endpoint rather than routing through a vendor. And it has to work with no outbound connection at all, which means bundled assets rather than fetched ones. All three are true here, which is why offline is a supported mode rather than a degraded one.

Setup

Install Ollama, pull a coding-oriented model, and point the app at the local endpoint. There is no account, no API key and no first-run call home. Disconnect the machine afterwards and behaviour is unchanged, which is the test that matters.

The quality tradeoff, stated plainly

Local open models are behind frontier hosted ones. The gap is largest on long multi-step reasoning and smallest on structured tasks like reading a repository and emitting a graph, which is why diagram generation holds up locally better than open-ended refactoring does. If your constraint permits a hosted model you will get better results; this exists for when it does not.

Mixing local and hosted

Roles can use different models, so a local model can handle everything that touches source while a hosted one handles work that does not. That is a middle position many organisations can actually approve, and it is more useful than the all-or-nothing framing the choice usually gets.

What changes when you go local

Local modelHosted model
Code leaves the machineNoYes, to your provider
API key requiredNoYes, yours
Marginal cost per turnZeroProvider's token pricing
Works with no networkYesNo
Quality on structured extractionGoodBetter
Quality on long reasoningNoticeably behindBetter
SpeedBounded by your hardwareBounded by the provider

Setup

shell
$ ollama pull qwen3-coder
$ ollama serve
# Point Lucid Train at http://localhost:11434
FAQ

Questions

Which local model should I use?

A coding-oriented one sized to your machine's memory. Larger is better where it fits; a model that swaps is unusable regardless of its benchmark scores.

Is any telemetry sent?

None unless you opt in. That is the default rather than a setting to find.

Can I use a self-hosted inference server?

Yes, anything OpenAI-compatible, including vLLM or a shared internal endpoint. Ollama is the easy path, not the only one.

Related