Text generation models

T3 Code writes some text for you: thread titles, commit messages, pull request titles and descriptions, branch names, and voice transcript cleanup. Pick the model for this in Settings → Threads → Text generation model. Settings → Source control and Settings → Voice can each pick a different one.

By default, a coding agent such as Codex or Claude writes this text, starting a session for each request. A text generator — a model API that only writes text — is usually faster and cheaper.

Add a text generator

Go to Settings → Providers → Text generators and choose Add generator:

• OpenAI or Anthropic — enter an API key.
• OpenRouter — enter an OpenRouter key to use any model it serves.
• Custom (OpenAI-compatible) — enter the base URL of any other endpoint, such as Ollama at http://localhost:11434/v1 or LM Studio. Local endpoints usually need no key. The environment's server makes the requests, so on a remote environment localhost means that machine, not yours.

Give it a name if you add more than one from the same service. Your key is stored on the environment, like other secrets, and can come from a secret provider instead of being typed in.

T3 Code lists the service's models to check the connection, then shows the model picker for each text task so you can switch any of them to the new generator, or keep what they use. Test next to a picker has the chosen model answer a tiny prompt, which proves the key works; some services list their models without checking it. Show models in the generator's menu lists its models and refreshes the list after the service adds some.

Each request sends the API what the task needs, such as the staged diff for a commit message.

Turning a generator off or removing it sends its work back to the environment's text generation model.

What text generators can't do

Review guides read the checkout over many steps, so they always use a coding agent. If your source control writer model is a text generator, review guides use Codex or Claude instead, and fail if neither is set up. To choose which, pick one in Settings → Source control → Review guide model.

A managed policy that limits providers also turns off text generators.