Models & ProvidersNEW
How Sophon connects to LLMs — bring-your-own-key, many providers, and flexible authentication.
Sophon is provider-agnostic and bring-your-own-key. It ships with no pre-configured providers — you connect the models you want, with your own credentials, and Sophon routes work to them. The model layer is part of the core engine and is fully decoupled from the orchestration pipeline, so adding a provider never changes how agents behave.
Sophon never bundles API keys or resells inference. Every token you spend is billed by the upstream provider on your own account, or covered by a subscription you already pay for.
What the model layer does
When the orchestration pipeline reaches a model call, it asks the model layer for a provider that matches the agent's routing rules, then streams the request and tokens back. Because the layer is provider-agnostic, the same agent can run on a frontier cloud model today and a local model tomorrow with only a config change. You can also override the model for a single conversation — the Dashboard chat inspector has a model picker, and the CLI has a /provider command — without touching your routing rules.
Sophon supports 30+ providers across four families:
| Family | Examples | Auth |
|---|---|---|
| Cloud LLM | Anthropic, OpenAI, Google Gemini, Azure OpenAI, AWS Bedrock, Mistral, Groq, DeepSeek, OpenRouter, xAI, Qwen, Zhipu | API key (some need extra fields) |
| Local / self-hosted | Ollama, LM Studio | None — set a base URL |
| OpenAI-compatible | Custom endpoint, Sophon Forge | API key + base URL |
| Subscription / CLI sign-in | Anthropic (Claude Pro/Max), OpenAI Codex (ChatGPT), GitHub Copilot, Claude Code | CLI credential sync or OAuth |
The full table, including required fields per provider, lives on Supported Providers.
Authentication types
Every provider declares one of four auth types:
- API key — the default for cloud providers. Paste a key; some providers also need extra fields (Azure OpenAI needs a resource endpoint and deployment name, Bedrock needs an access key, secret, and region, Cloudflare needs an account id and token).
- Subscription / CLI sync — Sophon reads tokens from an external CLI's credential store. For example, the Anthropic subscription provider reuses your
claude auth logincredentials from the Claude Code CLI, so a Claude Pro or Max plan powers Sophon with no separate key. - OAuth — Sophon runs the OAuth 2.1 + PKCE flow itself (for example, GitHub Copilot's device-flow sign-in) and refreshes tokens for you.
- None — local engines like Ollama and LM Studio need no credentials; you only provide a base URL.
Subscription and CLI providers report a per-token cost of $0 because they are flat-rate — useful to know when you set budgets.
Adding a provider
Via Dashboard
- Go to Settings > Models & Providers.
- Click Add Provider.
- Pick the provider and supply its credentials (API key, base URL, or sign-in).
- Test the connection.
- Set priority and budget limits.
Via CLI
sophon models addThis launches an interactive wizard that walks through the same steps and is handy for headless or scripted setups.
Where credentials are stored
Provider configurations are written to models.json under ~/.sophon/config/. The file is encrypted at rest — keys never sit in plaintext on disk. See Configuration for the full layout of the config directory.
In this section
- Supported Providers — the complete catalog, with required fields and notes for each provider.
- Model Catalog — the bundled registry of 20+ models with capabilities and pricing, per-session model selection, and custom overrides.
- Subscription & CLI Sign-In — power Sophon with Claude Pro/Max, ChatGPT (Codex), or GitHub Copilot instead of pay-per-token keys.
- Routing, Failover & Budgets — priority order, automatic failover between providers, and per-provider spend limits.
Related
- Configuration — first-run setup and the
~/.sophon/configfiles. - Extended Thinking & Reasoning — using models that expose a reasoning step.