Protocol and provider status¶

The model catalog and protocol implementations are independent. GetModels returns catalog entries. A request is routable only when the current process has registered an adapter for the model's Protocol. Use GetRunnableModels for runtime model lists.

Protocol status¶

Protocol	Catalog models	Status	Registration	Meaning
`openai-completions`	234	Implemented	`_ "github.com/ktsoator/or/llm/openai"`	OpenAI Chat Completions and compatible endpoints
`anthropic-messages`	73	Implemented	`_ "github.com/ktsoator/or/llm/anthropic"`	Anthropic Messages and compatible endpoints
`openai-responses`	55	Implemented	`_ "github.com/ktsoator/or/llm/openai"`	OpenAI Responses API and compatible gateways
`google-generative-ai`	20	Catalog only	None	No adapter
`mistral-conversations`	22	Catalog only	None	No adapter

Importing github.com/ktsoator/or/llm/all registers the three implemented protocols. It does not implement the two catalog-only protocols.

Catalog versus runtime state¶

models := llm.GetModels("openai") // returns cataloged openai-responses models
for _, model := range models {
    fmt.Println(model.ID, llm.SupportsProtocol(model.Protocol))
}

runnable := llm.GetRunnableModels("openai") // returns the runnable Responses models

GetModel and LookupModel only check catalog membership. They do not check adapter registration. Call SupportsProtocol before sending, or build selection UIs from GetRunnableModels.

Provider catalog¶

The current catalog contains the following provider IDs. Counts come from llm/catalog.generated.json.

Provider ID	Models	Protocol	Credential variables
`anthropic`	13	Anthropic Messages	`ANTHROPIC_API_KEY`
`cerebras`	3	OpenAI Completions	`CEREBRAS_API_KEY`
`deepseek`	4	OpenAI Completions	`DEEPSEEK_API_KEY`
`fireworks`	17	Anthropic Messages	`FIREWORKS_API_KEY`
`github-copilot`	20	Both implemented protocols	`COPILOT_GITHUB_TOKEN`
`google`	20	Google Generative AI	`GEMINI_API_KEY`
`groq`	6	OpenAI Completions	`GROQ_API_KEY`
`huggingface`	58	OpenAI Completions	`HF_TOKEN`
`kimi-coding`	4	Anthropic Messages	`KIMI_API_KEY`
`minimax`	7	Anthropic Messages	`MINIMAX_API_KEY`
`minimax-cn`	7	Anthropic Messages	`MINIMAX_CN_API_KEY`
`mistral`	22	Mistral Conversations	`MISTRAL_API_KEY`
`moonshotai`	10	OpenAI Completions	`MOONSHOT_API_KEY`
`moonshotai-cn`	10	OpenAI Completions	`MOONSHOT_API_KEY`
`nvidia`	58	OpenAI Completions	`NVIDIA_API_KEY`
`openai`	30	OpenAI Responses	`OPENAI_API_KEY`
`opencode`	56	All three implemented protocols	`OPENCODE_API_KEY`
`opencode-go`	18	All three implemented protocols	`OPENCODE_API_KEY`
`together`	18	OpenAI Completions	`TOGETHER_API_KEY`
`xai`	6	OpenAI Completions	`XAI_API_KEY`
`xiaomi`	3	OpenAI Completions	`XIAOMI_API_KEY` or `MIMO_API_KEY`
`xiaomi-token-plan-ams`	2	OpenAI Completions	`XIAOMI_TOKEN_PLAN_AMS_API_KEY`
`xiaomi-token-plan-cn`	2	OpenAI Completions	`XIAOMI_TOKEN_PLAN_CN_API_KEY`
`xiaomi-token-plan-sgp`	2	OpenAI Completions	`XIAOMI_TOKEN_PLAN_SGP_API_KEY`
`zai`	4	OpenAI Completions	`ZAI_API_KEY`
`zai-coding-cn`	4	OpenAI Completions	`ZAI_CODING_CN_API_KEY`

Built-in credential configuration is defined in llm/keys.go. At runtime, call APIKeyEnvVars(provider) to query the variable names for a provider.

Validation scope¶

Built-in tests cover the adapters for all three implemented protocols with local mock servers. The project does not run continuous live integration tests against every provider in the table. Therefore:

“implemented” means the built-in adapter handles the protocol's request and response formats;
“model is listed” means only that its metadata exists in the built-in model catalog; it is not a live-provider compatibility guarantee;
providers using non-standard fields may require Model.Compatibility or RewriteRequest;
before production use, validate authentication, streaming, tools, token usage, and error handling with an account for the target provider.

Model capabilities¶

Read capabilities from each Model:

Capability	Field or API
Image input	`slices.Contains(model.Input, llm.ModelInputImage)`
Reasoning	`model.Reasoning`, `SupportedThinkingLevels(model)`
Context window	`model.ContextWindow`
Maximum output	`model.MaxTokens`
Catalog price	`model.Cost`
Routable now	`SupportsProtocol(model.Protocol)`

The catalog is generated from external sources and embedded into the binary. Prices, model status, and limits can lag behind a provider. Usage.Cost is a catalog-priced estimate, not an invoice.