Skip to content

Models and keys ​

Agents need a model. There are two ways to have one, and a platform may offer both.

Your own keys. In Settings → Models, add a provider (Anthropic, or anything speaking the OpenAI API, including a model running on your own hardware) with your API key, then authorize the models you want to use. The key is encrypted at rest and never shown again, only as a hint like …abcd.

Models the platform offers. A platform administrator can publish models everyone can use, without anyone handling a key. They appear in the list marked as the platform's. The platform decides whether each one is offered to personal spaces, to organizations, or to everyone.

The platform may also limit what each person and each organization spends on its models per month. Your account page shows your limit and what you've used, and an organization's Settings → Budgets shows its own. Once it's used up, new answers on the platform's models are refused until the 1st (UTC). Models with your own keys are never limited by it.

Adding a provider ​

  1. Settings → Models → Add model provider, then the way it connects:
    • API key: pick the provider (Anthropic, OpenAI, xAI, Google Gemini, Mistral, Groq or OpenRouter) and paste its key. CrewHall knows their addresses.
    • Sign in: OpenRouter or your Grok subscription, in a personal space (below).
    • Local: Ollama, LM Studio or any OpenAI-compatible server (vLLM, LiteLLM, llama.cpp…). Give its address; a key only if your server asks for one. From Docker, localhost is the container itself: use host.docker.internal.
  2. Discover models asks the provider what it has, and doubles as a connection test.
  3. Authorize the models you want. Only authorized models can be chosen by agents.

Add model provider: an API key, signing in, or a model on your own hardware

Mark whether a model supports tools. Without that, agents on it cannot use tools, integrations, delegation or memory writing: they can only answer.

Mark whether it sees images. People can attach pictures in chat only when the agent's model does; for everyone else, an image upload is refused with a message naming the model. New Anthropic models start with it on; for other providers, check the model's documentation before turning it on.

Providers and the models authorized from them, with their prices

Signing in instead of pasting a key ​

In your personal space, two providers can be connected by signing in, under Add model provider → Sign in. Organizations use API keys.

OpenRouter. Sign in with OpenRouter takes you to OpenRouter, which creates a key for CrewHall and sends you back to pick models. OpenRouter serves models from many vendors (Claude, Grok, GPT, Gemini and more), paid from your OpenRouter credits. The key is yours: you can see, limit or revoke it in OpenRouter's settings. Signing in again replaces it.

Grok (SuperGrok or X Premium). Sign in with Grok shows a short code. Open xAI's page, check the code, and approve. Grok models then run on your subscription instead of paid API credits, and CrewHall renews the sign-in on its own. Two things to know:

  • Only you can run agents on it. If you share an agent that uses it, the people you share it with get a message saying the model is your subscription. Pick an API-key model for agents you share.
  • CrewHall signs in the way xAI's own Grok CLI does. xAI hasn't published a way for other apps to do this, so it may stop working without notice. If it does, sign in again, or use an xAI API key or OpenRouter instead.

Claude subscriptions (Pro, Max) can't be used this way: Anthropic allows them only in its own apps. Use an Anthropic API key, or Claude through OpenRouter.

Prices ​

Every model has an input and an output price, and CrewHall costs every call with it, which is what makes the usage page and budgets work. A call's cost is kept when it's made, so a later price change never rewrites past usage.

Public prices are followed. Every hour, CrewHall reads the public prices: LiteLLM's community price list, and OpenRouter's own list for OpenRouter models. Every model they know takes its public price, including one you typed by hand, once that public price moves. Hover over a price to see what it follows. Models no list knows (a local model, a custom endpoint) keep the price you set; set one, or their cost stays unknown.

  • A change applies at once. The space's owners and admins, and the editors of agents using the model, get a notification with the old and new price (by email too, under Budgets and prices in their notification settings).
  • A big jump, more than twice or less than half the current price, could be a mistake in the list. The current price stays, and the Models page asks a manager to take the new price or keep the current one. Keeping it isn't asked again until the list moves.

Prices are the list price in US dollars, before tax, at the provider's global rate. A negotiated discount or a regional endpoint costs something else; you can type it, but it's replaced the next time the public price changes.

A model with no price still works; its tokens are counted but its cost is not, and the usage page says how many tokens went unpriced.

Speech to text ​

A speech-to-text model lets people talk to agents instead of typing (see Talking instead of typing). It never answers as an agent: it writes down what people say. Authorize one like any model, on an OpenAI-compatible provider with an API key (OpenAI's whisper-1 or gpt-4o-mini-transcribe, xAI's grok-voice-transcribe-2.0, Groq's whisper-large-v3-turbo, Mistral's Voxtral…), and set its Kind to Speech to text. For Grok, that's a key from xAI's console: signing in with a SuperGrok subscription doesn't include it. Ids with whisper or transcribe in them are set to it already. Its price is per minute of audio, in the first price column.

The microphone shows wherever such a model is: in your space for your own, and in the spaces the platform offers one to. With none, there's no microphone.

Context window and limit per answer ​

Two more columns, both optional:

  • Context: how many tokens the model reads at most. With it, the chat shows how full a conversation is, and compacts it before it gets too full (long conversations). It's filled in from the public price list for the models it knows.
  • Limit / answer: the most tokens one answer may use on this model, all its calls together (each call sends the conversation so far again, so an answer that uses many tools adds up). Empty means no limit. At the limit, the agent answers with what it found so far and says it was cut short. Your space's monthly budgets still apply.

When a model disappears ​

Disable a provider, remove a key, or withdraw a model, and every agent using it shows the model as unavailable and stops being able to run until an editor picks another. Deleting a provider is refused outright while a published agent depends on it, rather than quietly breaking it.

CrewHall