# Models and API keys

AgentRun runs with your model credentials. Choose the provider in the console, then select a model that your chosen harness can use.

## Add a provider key

Open **Settings → Tool keys → Model keys** and save the key for your provider. The change applies without restarting the server. The launcher also reads keys already set in its environment.

| Provider | Environment variable | Starting point |
| --- | --- | --- |
| OpenRouter | `OPENROUTER_API_KEY` | A capped key gives you access to multiple model families. Start with Pi and the server's available default. |
| Anthropic | `ANTHROPIC_API_KEY` | Use an Anthropic model, including the native Claude Code harness configuration. |
| OpenAI | `OPENAI_API_KEY` | Use an OpenAI model, including the native Codex harness configuration. |

A Claude Code, Codex, or other coding-app login is separate from the credentials of the AgentRun server. Keep keys in the console or your secret manager; never include them in agent instructions, source control, or chat.

## Choose a model

An agent definition can set `model` to `provider/model-id`. With no model on the definition, it uses the server default. Set `AGENT_MODEL` when starting the server to choose that default explicitly.

With no explicit default, the server chooses from configured providers. Inspect `GET /v2/capabilities` rather than assuming which key or model won:

| Field | Meaning |
| --- | --- |
| `models.default` | The configured default model. |
| `models.ready` | Whether a revision using the default can run. |
| `models.needs` | Whether configuration is missing a key or a recognized model. |
| `models.priced` | Whether the server can estimate the default model's cost for a capped run. |
| `harnesses` | Installed harnesses, availability, and reasons for missing setup. |

The capability fields described here follow current repository source. An older installed server may expose fewer fields; its returned schema is authoritative.

## Gateways and workflow decisions

For an OpenAI Responses-compatible gateway, the server supports `AGENT_MODEL_BASE_URL`, `AGENT_MODEL_API_KEY`, and `AGENT_MODEL`. Gateway model IDs, harness routing, and pricing need to match that gateway. Use the [models reference](/docs/server-reference#models) before configuring one.

Workflow model nodes can use `fast`, `default`, and `strong` tiers. `AGENT_MODEL_FAST` and `AGENT_MODEL_STRONG` override those tiers. Typed decision nodes may use an OpenRouter-backed decision model or a separate decision provider such as Jev, depending on configuration. A plain agent task does not require Jev. See [Decision workflows](/docs/server-reference#decision-workflows) only when your workflow uses them.

## Control spend

Use a provider-side credit limit and set a run's `max_cost_usd`. An agent definition can supply a default cap; a run-specific cap overrides it. Caps depend on the server's pricing data and accounting mode, so they are not a substitute for the provider's billing limit.

An unknown price is not zero cost. A capped run using an unpriced model can be refused before it starts; configure the gateway price or use a priced model. For a paused run, inspect its pending budget decision in the console before continuing.

See [Spend cap](/docs/server-reference#spend-cap) for accounting modes, learning costs, and continuation behavior.