Custom API providers
Connect an API endpoint, discover its models, and reuse it in agent launch profiles.
Open Settings → Providers → Add provider. Give the provider a name, enter its API key, and add an endpoint with its protocol and base URL. Include the API prefix, such as https://api.example.com/v1. Remote endpoints require HTTPS; HTTP is supported for localhost.
Choose Fetch models to read the endpoint's model list. If discovery is unavailable, use Add model and enter the provider's model ID. Context windows, output limits, and reasoning support are optional: enter them only when the provider documents them. Without a known context window, atrium displays token usage without a percentage. Reasoning controls appear only for models marked as supporting them.
The first release supports these combinations:
| Adapter | Protocols |
| Codex | Responses |
| Claude Code | Anthropic Messages |
| OpenCode | Chat Completions, Responses, Anthropic Messages |
| Goose | Chat Completions, Anthropic Messages |
Compatible installed adapters are enabled by default. You can disable individual adapters in the provider editor. Pi and other adapters are not supported yet.
Claude Code gateway support is intended for supported Claude models. An endpoint implementing the Anthropic protocol does not make other models supported by Claude Code. See Anthropic's gateway documentation and legal and compliance guidance.
Launch and reuse
Provider models appear alongside built-in models in compatible launchers, labeled with the provider's name. Choose one before launching, then save it in a launch profile to reuse it for tasks and scheduled runs. A running chat can switch between models on the same provider endpoint; launch another pane to change providers.
Custom-provider panes show context usage, without subscription quota bars or account rotation. Authentication errors direct you to the provider's API key settings. Removing a provider or losing its saved key prevents launches that reference it; atrium does not silently use a subscription account.
Keys belong to the execution host
Keys are stored in macOS Keychain, Windows Credential Manager, or the Linux login keyring through secret-tool from libsecret. Linux hosts need an unlocked keyring and secret-tool installed. A locked or unavailable credential store produces an error; there is no plaintext fallback.
Launch profiles and pane snapshots store provider references. Remote and mobile launches resolve the provider and its key on the host running the agent. Configure the provider on that host before using it there.
Secret-valued environment settings are also stored as credential references. Existing settings migrate at startup, and pane profiles migrate before launch. Previously created backups may still contain old values; migration does not rewrite backup history. Runtime terminal environment files are excluded from atrium's snapshot repository and use owner-only permissions on macOS and Linux.
Provider telemetry records the adapter and protocol, without the endpoint URL, provider label, model ID, or API key.