Set up
Models
A model connection is a saved API key or endpoint for a model provider; agents can only use the models an admin allows on it.
Every agent needs a text model to answer. Agents can also have an image model, to generate images, and a decision model, for fast structured choices. Owners and admins connect providers and allow models in Settings → Models; anyone who can edit an agent picks from the allowed models.
Supported providers
| Provider | Connection needs | Text | Image | Decision |
|---|---|---|---|---|
| Anthropic | API key | Yes | Yes | |
| OpenAI | API key | Yes | Yes | Yes |
| OpenRouter | API key | Yes | Yes | |
| Fireworks | API key | Yes | ||
| Together AI | API key | Yes | Yes | |
| Ramp Router | API key | Yes | ||
| TypeSafe AI | API key | Yes | ||
| OpenAI compatible | Base URL, optional API key | Yes | Yes |
A provider's column says what kind of model it can supply. Whether a particular model supports images or decisions comes from the provider's model list.
Connect a provider
- Open Settings → Models and choose a provider under Add a provider.
- Name the connection and paste the API key. Each provider's form links to the page where you create one.
- Choose Connect.
Adapt saves the connection, then lists the provider's models with the key. The connection shows one of these states:
- Connected: the provider accepted the key and returned its model list.
- Key rejected: the provider refused the key. The connection is saved anyway; edit it with a working key.
- Not tested: Adapt couldn't check, for example because the provider was unreachable.
- Disabled: an admin turned the connection off. Agents can't use it until it's enabled again.
A connected key proves the provider answered a model-list request. It doesn't prove the account has credit or can run every model. Choose Test key from the connection's menu to check again later.
When the first connection to a provider fills an empty workspace default, Adapt allows that provider's recommended model and makes it the default for new agents. Today that's claude-opus-5-5 (text) and claude-haiku-5-5 (decision) for Anthropic, and gpt-6.1-sol (text), gpt-image-2.5-flare (image), and gpt-6-luna (decision) for OpenAI, when the provider lists them for your key. Existing defaults are never replaced.
Allow models
On a connection in Settings → Models, choose Add model, search the provider's list, and pick one or more models. Only allowed models appear in agent pickers, and a run fails with an error if its agent's model is no longer allowed.
To remove a model, choose the remove button next to it. Adapt won't remove a model that agents still use; it names those agents so you can switch them first. Removing the workspace default clears it, and new agents then need a model picked by hand.
Choose a model for agents
Set the starting models for new agents in Settings → Defaults → Models. Change one agent's models in that agent's Settings tab, under Models. Each picker lists the allowed models of enabled connections:
- Model: the text model the agent thinks and calls tools with. Required.
- Image model: lets the agent generate images and attach them to its reply.
- Decision model: lets the agent make structured choices, such as yes or no, a pick from a list, or a score. It also decides whether a Slack channel message needs a reply; see Slack.
Reasoning effort sets how much the model thinks before answering: Default, Low, Medium, or High, plus Max for Anthropic. Default leaves it to the provider. If the model doesn't support the effort you set, the run fails with an error instead of ignoring it.
Changing Agent defaults doesn't change existing agents. Model changes apply from the next run; a run in progress keeps the model it started with.
Use a local or self-hosted model
Choose OpenAI compatible to connect a local inference server or another gateway, such as Ollama, LM Studio, or vLLM.
- API base URL: the endpoint's root, usually ending in
/v1, for examplehttp://localhost:11434/v1. - API key: optional. Leave it blank for an unauthenticated local server.
The Adapt server makes these requests, not your browser, so the URL must resolve from the machine running Adapt. If Adapt runs in a container, localhost means that container. The endpoint must serve:
GET /models, which Adapt reads to list the models you can allow.- Chat completions with tool calling. Agents can't use tools without it.
Large tool sets
Some models can search for tools instead of receiving every tool definition up front: Anthropic's Claude Opus, Sonnet, and Haiku 4.5 and later, every Claude Fable and Mythos model, and OpenAI GPT models from GPT-5.4, except nano and pro models. With these models, a connection whose tool definitions reach 20,000 characters stays out of the prompt until the model searches for one of its tools. Built-in tools and smaller connections are always sent in full. Other models receive every tool.
If a provider rejects one tool's definition, the run continues without that tool and the model is told why.
Limits
These run limits are set in Agent defaults and copied into each new agent:
- Max output tokens: 4,096 by default, at least 256.
- Context budget: 60,000 characters by default, at least 12,000. When history grows past it, older turns are summarized. It approximates context size in characters, so set it for your model's actual context window.
Model API keys stay on the Adapt server. They're never placed in sandbox environments.