Features / Models

Your models, your call

Start on small local models through Ollama. Add more power when you want it, without editing a file or restarting anything.

Managed from the browser

Settings → Models

Admins add providers and keys, test them, and pick the default model. Keys saved here are encrypted and never shown again.

Only your own Ollama is on at first. Cloud providers wait until an admin turns them on.

Providers

Who gets what

Every model, or a list: Offer everything your providers have, or only the models you choose, with labels for the pickers.
Admins only: Chosen models can be limited to admins. Members don't see them, and the web chat, agents, tasks and Studio refuse them.
Answer modes: Fast, Auto or Thinking sets how much a model thinks before it answers, next to the model picker.
Each chat its own: Picking a model in a chat changes only that chat, and it stays with it on your other devices too.
Prices: Set what paid models cost, or fill them in from a public price list to check, and Usage shows what replies cost.
Context: Set how much a chat may fill before it's summarized, and how much Ollama keeps in memory.

Made for small models

Models get tool calls wrong now and then. labb makes each reply's calls valid before they run, in the terminal and the web chat alike.

Repair: Close names, misspelled arguments and broken JSON are fixed.
Constrained retry: With Ollama, a call that still doesn't fit is asked for again, constrained to the tool's schema.
Strict mode: With OpenAI, Anthropic and compatible servers, the provider constrains the arguments itself.
A smaller prompt: The labb local engine offers fewer tools with shorter descriptions, and starts at about 1,900 tokens instead of about 5,100.
Measure it: Usage shows the share of calls that were right as written, and labb eval tools scores your models on 35 prompts. labb eval long runs longer jobs for real and checks the files.

More features