llmgwA provider instance is one configured ID, runtime type, base URL, and settings. A provider connection is an encrypted credential envelope owned by a human or system principal.
Registry matrix
| Integration | Protocol | Authentication | Scope / caveat |
|---|---|---|---|
| GitHub Copilot | OpenAI-shaped | Device OAuth | personal Grey-area gateway use; disabled by default. |
| Claude Code | Anthropic | Gateway key | Client-only setup, not an upstream provider. |
| OpenAI Codex | Responses | Device OAuth + PKCE | experimental Personal; client ID optional. |
| OpenAI | OpenAI | API key | System or personal. |
| Anthropic | Anthropic | API key or setup token | Native Messages and token count; encrypted setup-token storage. |
| Google Gemini | OpenAI-compatible | API key | Google's compatibility endpoint. |
| Google AI Studio | Gemini | API key | Chat with streaming and tools, image, video. |
| Vertex AI (Agent Platform) | Gemini/Vertex | API key or service account | Project/location scope; Cloud billing. |
| Google Antigravity | Cloud Code Assist v1internal | Browser OAuth | experimental Personal; undocumented API and operator-provided OAuth client required. |
| OpenRouter | OpenAI-compatible | API key | Multi-vendor aggregator. |
| Groq | OpenAI-compatible | API key | Hosted low-latency inference. |
| Mistral AI | OpenAI-compatible | API key | Hosted Mistral models. |
| Cerebras | OpenAI-compatible | API key | Hosted inference. |
| DeepSeek | OpenAI-compatible | API key | DeepSeek's own models. |
| Moonshot AI (Kimi) | OpenAI-compatible | API key | International platform. |
| Moonshot AI (Kimi), China | OpenAI-compatible | API key | China platform; separate keys. |
| Zhipu BigModel | OpenAI-compatible | API key | GLM models in China; no model discovery. |
| Z.ai | OpenAI-compatible | API key | Zhipu's international platform; no model discovery. |
| Alibaba Cloud Model Studio (Qwen) | OpenAI-compatible | API key | Singapore region; region-bound keys; no model discovery. |
| Alibaba Cloud Model Studio (Qwen), China | OpenAI-compatible | API key | Beijing region; region-bound keys; no model discovery. |
| Alibaba Cloud Model Studio (Qwen), US | OpenAI-compatible | API key | Virginia region; region-bound keys; no model discovery. |
| Alibaba Cloud Coding Plan | OpenAI-compatible | Coding Plan key | personal Alibaba Cloud allows plan keys only in interactive coding tools. |
| Alibaba Cloud Coding Plan (Anthropic API) | Anthropic | Coding Plan key | personal The same plan and terms over Messages. |
| Alibaba Cloud Coding Plan, China | OpenAI-compatible | Coding Plan key | personal China site; the same terms. |
| Alibaba Cloud Coding Plan, China (Anthropic API) | Anthropic | Coding Plan key | personal China site over Messages; the same terms. |
| Volcengine Ark | OpenAI-compatible | API key | Doubao and other models in China; no model discovery. |
| SiliconFlow | OpenAI-compatible | API key | Open models; international site. |
| SiliconFlow, China | OpenAI-compatible | API key | China site; separate keys. |
| NVIDIA NIM | OpenAI-compatible | API key | terms Hosted catalog, free for prototyping. |
| Venice | OpenAI-compatible | API key | Private inference of open models. |
| NEAR AI Cloud | OpenAI-compatible | API key | Models in confidential computing. |
| MiniMax | OpenAI-compatible | API key | International platform. |
| MiniMax, China | OpenAI-compatible | API key | China platform; separate keys. |
| LongCat | OpenAI-compatible | API key | Meituan's LongCat models. |
| ModelScope | OpenAI-compatible | Access token | terms Free for non-commercial use within rate limits. |
| ModelScope, China | OpenAI-compatible | Access token | terms China site; the same terms. |
| Novita AI | OpenAI-compatible | API key | Open-model aggregator. |
| Avian | OpenAI-compatible | API key | Hosted open-model inference. |
| Vivgrid | OpenAI-compatible | API key | Model selected server-side; no model discovery. |
| ShengsuanYun | OpenAI-compatible | API key | Multi-vendor model router. |
| Amazon Bedrock | OpenAI-compatible | Bearer API key | Region required; no IAM/SigV4. |
| Azure OpenAI | Chat | Azure api-key | Model ID is deployment name; legacy discovery version. |
| Ollama | Ollama | None | local Native local discovery/chat. |
| LocalAI | OpenAI-compatible | Optional API key | local Audio capabilities may be inferred. |
| LiteLLM | OpenAI-compatible | Optional API key | local Self-hosted LiteLLM proxy. |
| vLLM | OpenAI-compatible | Optional API key | local Self-hosted vLLM server. |
| LM Studio | OpenAI-compatible | Optional API key | local LM Studio's local server. |
| Edge TTS | Speech | None by default | unofficial Public Edge read-aloud service. |
| ElevenLabs | Native audio | API key | System or personal. Direct speech-to-text and text-to-speech. |
| Xiaomi MiMo | Chat Completions audio | API key | System or personal. Direct text-to-speech through MiMo's audio contract. |
| Custom OpenAI-compatible | OpenAI-compatible | Optional API key | Actual surfaces depend on upstream. |
| OpenCode Zen | OpenAI-compatible | None by default; optional API key | System or personal. Anonymous multi-model endpoint; availability varies by model. |
| Kilo Code | OpenAI-compatible | None by default; optional API key | System or personal. Anonymous multi-model endpoint with auto/free routing. |
| LLM7.io | OpenAI-compatible | None by default | System or personal. Anonymous turbo-tier chat models. |
| OVH AI Endpoints | OpenAI-compatible | None by default | System or personal. Anonymous per-IP tier with strict rate limits; NVR voices on speech. |
| Pollinations.ai | OpenAI-compatible | None by default | System or personal. Anonymous text inference with provider-specific paths. |
| Custom Anthropic-compatible | Anthropic | API key | Requires Messages-compatible upstream. |
A row with no model discovery offers no model list to sync; name its models in exact targets. litellm also remains an advanced runtime type, beside the LiteLLM card.
github_copilot, openai_codex, google_antigravity, edge_tts, opencode_zen_anonymous, and anthropic_setup_token provider types are served by an optional, separately distributed companion daemon, which an administrator adds as an extension, with its address and shared secret, on the console's Providers page. Without one, those provider types are unavailable.Credential resolution
Generic API-key providers resolve the human's private default connection, then a system connection, then the legacy config/secret seed. Generic system fallback is not project-binding-gated today.
Copilot is stricter: humans use their own OAuth; services require an exact active project/provider/service binding to a gateway-owned credential. Codex OAuth remains human-private.
secrets.json for compatibility, so protect all state and backups.Four Google paths
Google Gemini uses the OpenAI-compatible endpoint. Google AI Studio uses native Gemini, with streaming chat and function tools, and supports image/video. Vertex AI uses native Gemini/Vertex, with the same chat, in a project and location scope. Experimental Google Antigravity uses owner-private browser OAuth and undocumented Cloud Code Assist v1internal APIs for exact catalog discovery and non-streaming chat. IDs and availability differ by API and location.
Local detection
The console can probe common addresses for Ollama, LocalAI, LM Studio, vLLM, Jan, text-generation web UI, and llama.cpp, including host.docker.internal. Detection reports candidates; the operator adds them. Silent auto-add is opt-in.
Anonymous provider automation
Set LLMGW_ANONYMOUS_PROVIDER_AUTOMATION=true to opt into the reviewed no-key profiles for OpenCode Zen, Kilo Code, LLM7.io, OVH AI Endpoints, and Pollinations.ai. The gateway adds missing default instances, filters their free catalogs, and runs one minimal verification request per provider at startup, limited by a durable 24-hour claim. It creates no credential and does not overwrite customized, credentialed, disabled, or existing provider instances.
Lifecycle evidence
- Check reachability: catalog endpoint answered.
- Sync catalog: model/capability rows persisted.
- Test completion: a real inference request succeeded.
- Clear cache and retry: local cache reset only; cannot repair credentials or entitlement.
Provider risk boundaries
Respect subscription terms. No Claude personal OAuth, browser-cookie extraction, stealth session reuse, or official-client impersonation is implemented. Google Antigravity relies on an undocumented API; Edge TTS and Azure discovery can also change independently of this project.