Skip to main content
A model provider supplies inference. A conversation connector receives and sends messages in Slack, Telegram, SMS or your application. You can change an agent’s model provider without replacing its conversation connections.

​
Supported providers

See OpenAI and self-hosted inference for endpoint configuration, keyless local servers, supported request features and limitations. A custom OpenAI endpoint applies to all agents using that adapter in the deployment; it is not a separate URL per agent. There is no implicit fallback from an unavailable local server to a cloud provider.

​
Credentials and instance defaults

An owner can configure Instance model defaults. New agents inherit the configured provider, its model assignments and its billing credential. An agent can instead choose a provider and store its own encrypted key, or use that provider’s instance key. The Models page shows the effective configuration. Switching providers does not transfer one provider’s key to another. Model keys are usage-payer credentials. Inheritance does not share an agent’s Slack identity, GitHub account, tool credentials, conversations or memory. OPENROUTER_API_KEY is a legacy process fallback, not the normal provisioning path. It applies only on a self-hosted (local authentication) deployment, only to agents that inherit instance defaults, and only until instance model defaults are saved; hosted deployments never use it. An unreadable saved credential must be repaired; it is not silently replaced with another account’s credential — not the instance key, not the process fallback, and not a managed account’s key that the instance no longer uses. When an instance stops being managed by a hosting control plane, its managed key records are deleted rather than kept as a fallback.

​
Model roles

Structured runtime decisions also have a separate configurable decision-provider path. See automatic model selection for routing, provider readiness and the relationship between a selected model and the primary fallback. Choosing a primary model does not enable every optional runtime feature.

​
Automatic model selection

Formation can choose a model for a run from the connected provider’s catalog under its routing policy. New agents start with the installation’s default, which is on unless an owner turns it off. The primary model is normally the fallback and the price ceiling (a family restriction that excludes it substitutes the best eligible model in that family), and turns use the primary until it is a concrete model the catalog lists. The Models page exposes the agent’s routing configuration; operators can inspect route decisions. Selection is bounded by capability and catalog evidence. Sparse catalogs can conservatively keep the primary model. Follow Automatic model selection for supported modes, configuration and limits.

​
Tools and failure handling

Model adapters translate the provider-neutral message and tool contract into the provider’s protocol. Formation decides which tools are available, validates their arguments, executes them and records their observations. A model request alone is not evidence that an action happened. OpenRouter can offer the same granted tools through a strict JSON decision protocol when native tool support is unavailable. Invalid output is rejected; a bounded repair can be attempted. Ramp Router and OpenAI use native tool calls and do not provide that text-protocol fallback. Runs record model calls, requested and resolved models, provider request IDs, finish reasons and provider-reported token usage. Retry and fallback behavior is bounded by the run’s budget and recovery policy. Tool rounds are budgeted per run; the short conversation profile is not a universal limit on coding work. There is no eight-call limit on a single batch: larger batches are validated before any call runs, then executed sequentially within the run’s allowance. If a batch exceeds the remaining tool-call budget, none of it runs. Formation records the refusal and tells the model how many calls remain so it can retry a smaller batch. Rejected calls do not spend the tool-call allowance; model iterations still bound recovery attempts. When the allowance is exhausted, the run closes its tools and produces a final status. See Agent harness, egress reliability and observability. The implemented adapter contract is in internal/model.