An engine is the agent that actually writes the code. Myrax does not ship one; it drives whichever you already have, on your machine. Four are detected today, and the choice is stored per project so a build cannot silently switch engine between the plan and the repair pass.
The four engines
| Engine | Detected as | What it needs |
|---|---|---|
| Claude Code | claude --version | Installed and signed in, or an Anthropic key saved in Providers. |
| Codex | codex --version | Installed and signed in, or an OpenAI key saved in Providers. |
| OpenCode | opencode --version | Installed, with its own provider credentials. Works with OpenRouter, so a free model id can run a build at no cost. |
| Ollama | ollama --version or a reachable local API | A downloaded model. Ollama cannot edit files, so the reply is parsed from a strict file-block format instead. |
Installed is not the same as ready
The dashboard reports two separate things per engine. installed means the command answered. ready means it also has a credential it can use. A machine can show Codex: installed and still be unable to build, and the build card says so directly: “needs a sign-in”, with the reason the agent reported.
This distinction is deliberate. An earlier version queued a build whenever the command existed, so the failure surfaced ten minutes later as a failed run instead of immediately as a missing credential.
How the engine is chosen
When you start a build you can name an engine and a model. If you do not, the server picks from what the target computer reports: an engine that is ready first, then one that is merely installed, then the historical default. The composer preselects the same way, so what you see matches what runs.
You can change it later with the engine picker on the build. A running stage refuses the change (engine-locked-while-running) so a single build cannot half-run on two engines.
Pin the model, especially on OpenCode
If you leave the model empty, OpenCode chooses its own default, and that default may not be one you want or can afford. On an OpenRouter account it has defaulted to an image model, which fails in seconds with a billing message about image output. Set the model on the project; the value is passed as --model and recorded with the build.
Local models
Ollama is a chat endpoint, not an agent that edits files. For builds, the reply is parsed from a strict block format:
<<<CALESTIX_FILE path="index.html">>> ...complete file contents... <<<END>>>
Write index.html first: the HTML is what makes the page exist, and a stylesheet without markup is an empty build. Small local models often run out of room mid-file, so a build that ends in only CSS is a truncation, not a mystery.
Provider keys and the Providers page: Providers and keys.

