Providers
Omegon treats provider authentication as a runtime boundary, not just a login convenience. The active provider affects routing, allowed entry points, quota semantics, and what the operator sees in the engine footer.
Provider authentication
The running contribution registry is authoritative for available providers. The examples below describe common authentication postures rather than an exhaustive provider or model inventory. A declared provider is not necessarily executable: route resolution still needs an executable contribution, accepted credentials or a declared local posture, and a constructible bridge.
| Provider / mode | Configure | Allowed | Restrictions |
|---|---|---|---|
| Anthropic/Claude API key | omegon secret set ANTHROPIC_API_KEY --stdin | All Omegon entry points, including TUI, prompts, smoke, and cleave. | Subject to Anthropic API terms, quota, and model limits. |
| Anthropic/Claude OAuth / subscription | omegon auth login anthropic or /auth login anthropic | Interactive TUI sessions; headless Anthropic runs with an explicit operator-risk warning. | Omegon does not treat subscription OAuth as equivalent to an API key. If you want policy-clean prompts, smoke tests, cleave, or unattended automation, useANTHROPIC_API_KEY instead. |
| OpenAI API | omegon secret set OPENAI_API_KEY --stdin | All supported Omegon entry points. | Subject to OpenAI billing, quota, and model limits. |
| OpenAI/Codex OAuth | omegon auth login openai-codex or /auth login openai-codex | Supported Codex-backed Omegon flows, including interactive work and supported automation. | The engine footer may show a selected model such as gpt-5.4 while the limit row shows an upstream bucket name such as GPT-5.3-Codex-Spark. That limit row is provider telemetry, not a second model selector. |
| Ollama (Local) | export OLLAMA_HOST="http://127.0.0.1:11434" | All supported local-model workflows. | Local inference only; no external account auth. |
| Ollama Cloud | omegon auth login ollama-cloud, /auth login ollama-cloud, or /secrets set OLLAMA_API_KEY | All supported hosted-Ollama workflows. | Separate hosted provider from local Ollama. Uses OLLAMA_API_KEY against Ollama's cloud API. |
| OpenAI-compatible API key providers | /auth login openrouter, /auth login groq, /auth login xai, /auth login mistral, /auth login cerebras, /auth login opencode-go, or /auth login perplexity | All supported key-backed workflows. | Each provider has its own billing, quota, model catalog, and upstream error semantics. |
| Google Gemini | /auth login google or GOOGLE_API_KEY / GEMINI_API_KEY | All supported key-backed workflows. | Gemini models through Google's public API surface. |
| Google Antigravity | /auth login google-antigravity | Authentication and inventory declaration only; the current provider contribution is non-executable. | A credential does not make this route dispatchable. Use the executable Google API-key route where appropriate. |
| GitHub Copilot | /auth login github-copilot | Interactive use with a Copilot-specific credential. | GitHub Copilot inference. Generic GitHub CLI and GITHUB_TOKEN credentials are diagnostic only and are not used for runtime inference. |
OpenRouter Ox Alpha preview
The embedded OpenRouter inventory and curated model menu includeopenrouter:stealth/ox-alpha. OpenRouter describes Ox Alpha as a free stealth preview for coding, sustained agentic work, and complex reasoning, with a 1,048,576-token input context and 131,072-token output limit. It accepts text, image, and video input, returns text, and supports function calling.
The underlying developer and operator are intentionally anonymous during the preview. Although the model listing currently says retained content is not used for training, OpenRouter's governing Stealth Program EULA says user content may be collected, shared, retained, and licensed for Stealth Model training, evaluation, and improvement. Do not send sensitive or restricted data. Treat this as a distinct broker route rather than inferred equivalence with another model family.
Route state and remediation
Omegon keeps the operator-selected route separate from the model bridge that is actually serving turns. That distinction matters when credentials expire, a rebuilt binary discovers external OAuth state, or a profile permits an explicit fallback provider.
- Selected route is the provider/model the operator or profile requested.
- Serving route is the authenticated provider/model that receives the request.
- Interactive fallback is credential fallback at startup, scoped by the ordered
fallbackProviders = ["provider-id"]list. The selected contribution must also declare directed, model-family-bounded compatibility with the serving provider. - Disconnected means no valid credential or allowed fallback is available; Omegon surfaces remediation instead of pretending another provider is active.
An empty fallbackProviders list does not authorize substitution. Compatibility is directed and non-transitive: if provider A declares B, that does not authorize B to use A or A to reach a provider declared only by B. Sessionless callers do not inherit the interactive list. Ordinary sessionless resolution may follow only fallback compatibility declared by the selected contribution; exact resolution probes only the selected provider.
Use /auth status when startup or login looks wrong. It reports route state, credential diagnostics, and the next remediation step. The TUI footer also preserves route warnings so a fallback or disconnected startup is still visible after the immediate login/status message scrolls away.
Authentication and credential evidence
A provider contribution declares an authentication class such as API key, OAuth, API-key-or-OAuth, OAuth token exchange, credentialless local, or optional-key local. This is a compatibility constraint, not a secret and not proof that a usable credential exists.
Before dispatch, the route lease records a bounded credential-source class such as environment, stored, external, or secrets manager. When a more specific source is unavailable, it may record the contribution's authentication class. This evidence never contains secret material and does not promise one exhaustive credential lookup order. Provider identity, selected route, serving route, and upstream limit telemetry remain distinct fields.
Retry and fallback are different
A provider-request retry repeats request handling on the captured serving route. Fallback resolves a different declared-compatible serving provider while retaining both selected and serving route identities and a bounded reason. Neither state authorizes replay of a tool or privileged invocation whose completion became unknown after owner handoff.
Anthropic subscription boundary
Anthropic/Claude subscription auth is the one credential class with an explicit operator-risk boundary. Interactive TUI use is allowed. Headless Anthropic paths warn and proceed rather than silently switching providers or pretending subscription OAuth has API-key automation semantics.
The consumer terms page is here:anthropic.com/legal/consumer-terms
Practical guidance
- If you want to work in the TUI by hand, Anthropic/Claude OAuth is a valid path.
- If you want unattended or scripted automation with clean provider terms, configure
ANTHROPIC_API_KEY. - If you want GPT-family routing without API-key setup, try
omegon auth login openai-codex. - If you want local-only inference, use
Ollama (Local). - If you want hosted Ollama without running a local daemon, use
Ollama Cloudvia/auth loginor/secrets. - If you want predictable automation with commercial API terms, prefer API-key-backed providers.
How this fits the rest of Omegon
The provider layer feeds the TUI, slash commands, smoke paths, and cleave orchestration. That means auth mode affects not just model access, but which entry points Omegon can honestly and safely expose.