Provider choices and model presets
src/providers.ts is the shared terminal and desktop provider registry. Its snapshot date is 2026-10-14. Each entry supplies the SDK base URL, protocol and either endpoint discovery or released preset ids. The provider choice is a form choice, not saved account metadata; saved credentials and auth keep their format. Regional and account-specific URLs can be edited before saving.
Endpoint discovery is used for OpenAI, Anthropic, Groq, Gemini’s OpenAI compatibility root, xAI, Mistral, Moonshot, MiniMax’s two compatible roots, StepFun, SiliconFlow, Qianfan, DeepSeek’s OpenAI root and Together. These entries carry no model presets. Discovery failures stay failures; the user may type a model id. See models for paths and pagination.
The following providers use a small released selection of ids. This is not a complete directory or proof that the entered key can access them. Presets are chosen explicitly before discovery; no model-list request is made in that flow. Searching always permits a different id.
| Provider | Preset evidence and base URL evidence |
|---|---|
| Zhipu GLM | Model overview, HTTP API, Anthropic compatibility. GLM ids are also present in the CC Switch 2026-10-13 preset snapshot; Circle uses the documented /api/paas/v4 OpenAI base. |
| Alibaba Bailian | Anthropic Messages supplies the supported ids and says /v1/models returns 404. Web search documents the Messages-route web_search tool and its required client identity in system, and the workspace *.maas.aliyuncs.com/apps/anthropic addresses matched as aliases. Base URL supplies the OpenAI compatibility base. The shared Beijing domain is still supported; workspace-specific domains can be entered instead. |
| Volcengine Ark | Model list, context examples, third-party tools. An account may require its own endpoint id; the model remains editable. |
| DeepSeek Anthropic | Anthropic compatibility and model list. The OpenAI root has an online list; the Anthropic base uses presets. |
| Tencent Hunyuan | OpenAI compatibility and model overview. These are Hunyuan ids, not TokenHub ids. |
| SenseNova | OpenAI compatibility documents /compatible-mode/v2 and the static model table. |
| Cohere | Compatibility API supplies command-a-plus-05-2026 and /compatibility/v1. |
| Fireworks | OpenAI compatibility, Cherry Studio provider snapshot. The account catalog API lives outside the inference base. |
| Amazon Bedrock | Anthropic Messages API supplies the API-key route, us.anthropic.claude-sonnet-5 and global.anthropic.claude-sonnet-5. The region in the default us-east-1 base and the inference-profile id can be edited. |
| Azure OpenAI / Microsoft Foundry Anthropic | Azure v1, Foundry Claude. Neither entry invents an account URL or deployment name: both fields must be entered. |
The ids are maintained in source and change with Circle releases. Updating this registry does not update model prices, context windows or capabilities; those remain the separate ModelCatalog responsibility. Presets never populate the runtime’s discovered model cache. /models, CLI --list-models and desktop refresh still request the saved endpoint’s actual list.
Vertex uses project/publisher prediction resources and dedicated authentication, so Circle does not offer a misleading generic Anthropic preset for it. Bedrock’s SigV4 integration is also outside this registry. The Bedrock preset uses only the compatible API-key Messages route; it does not sign AWS requests or manage temporary tokens.