{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"name": "llmrust",
"version": "0.1.1",
"description": "Machine-readable capability metadata for llmrust. AI agents can use this to discover what features each provider supports without reading source code. Note: 'support' means llmrust maps the field or protocol; actual upstream model support may vary.",
"protocol_version": "1",
"providers": {
"openai": {
"provider_key": "openai",
"description": "OpenAI GPT models (gpt-4o, gpt-4o-mini, etc.)",
"protocol": "openai-compatible",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": true,
"json_schema": true,
"logprobs": true,
"image_input": true,
"system_message": true,
"multi_turn": true,
"env_vars": ["OPENAI_API_KEY", "LLMRUST_OPENAI_KEY"],
"default_base_url": "https://api.openai.com/v1",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop", "seed", "presence_penalty", "frequency_penalty", "logprobs", "top_logprobs", "response_format", "n", "parallel_tool_calls", "service_tier", "store", "metadata", "user"]
},
"deepseek": {
"provider_key": "deepseek",
"description": "DeepSeek models (deepseek-chat, deepseek-coder)",
"protocol": "openai-compatible",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": true,
"json_schema": false,
"logprobs": false,
"image_input": false,
"system_message": true,
"multi_turn": true,
"env_vars": ["DEEPSEEK_API_KEY", "LLMRUST_DEEPSEEK_KEY"],
"default_base_url": "https://api.deepseek.com/v1",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop", "response_format"]
},
"moonshot": {
"provider_key": "moonshot",
"description": "Moonshot / Kimi models (moonshot-v1-8k, kimi-latest)",
"protocol": "openai-compatible",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": true,
"json_schema": false,
"logprobs": false,
"image_input": false,
"system_message": true,
"multi_turn": true,
"env_vars": ["MOONSHOT_API_KEY", "LLMRUST_MOONSHOT_KEY"],
"default_base_url": "https://api.moonshot.cn/v1",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop"]
},
"openrouter": {
"provider_key": "openrouter",
"description": "OpenRouter (access any model through OpenRouter)",
"protocol": "openai-compatible",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": true,
"json_schema": true,
"logprobs": true,
"image_input": true,
"system_message": true,
"multi_turn": true,
"env_vars": ["OPENROUTER_API_KEY", "LLMRUST_OPENROUTER_KEY"],
"default_base_url": "https://openrouter.ai/api/v1",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop", "seed", "presence_penalty", "frequency_penalty", "response_format", "n"]
},
"anthropic": {
"provider_key": "anthropic",
"description": "Anthropic Claude models (claude-sonnet-4-20250514, claude-3-5-sonnet, etc.)",
"protocol": "anthropic-native",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": false,
"json_schema": false,
"logprobs": false,
"image_input": true,
"system_message": true,
"multi_turn": true,
"env_vars": ["ANTHROPIC_API_KEY", "LLMRUST_ANTHROPIC_KEY"],
"default_base_url": "https://api.anthropic.com/v1",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop_sequences"],
"notes": ["Uses native Anthropic Messages API (not OpenAI-compatible)", "Images via base64 in content blocks", "Tool calling via tool_use/tool_result content blocks"]
},
"google": {
"provider_key": "google",
"description": "Google Gemini models (gemini-2.0-flash, gemini-1.5-pro)",
"protocol": "gemini-native",
"chat": true,
"stream": true,
"tool_calling": true,
"tool_calling_stream": true,
"json_mode": true,
"json_schema": true,
"logprobs": true,
"image_input": true,
"system_message": true,
"multi_turn": true,
"env_vars": ["GOOGLE_API_KEY", "LLMRUST_GOOGLE_KEY"],
"default_base_url": "https://generativelanguage.googleapis.com/v1beta",
"sampling_params": ["temperature", "max_tokens", "top_p", "stop", "logprobs", "top_logprobs", "response_format"],
"notes": ["Remote http(s) image URLs are skipped. Use data: URLs only.", "API key sent via x-goog-api-key header"]
},
"ollama": {
"provider_key": "ollama",
"description": "Ollama local models (llama3.2, qwen2.5, any local model)",
"protocol": "ollama-native",
"chat": true,
"stream": true,
"tool_calling": false,
"tool_calling_stream": false,
"json_mode": false,
"json_schema": false,
"logprobs": false,
"image_input": false,
"system_message": true,
"multi_turn": true,
"env_vars": ["OLLAMA_HOST", "LLMRUST_OLLAMA_HOST"],
"default_base_url": "http://localhost:11434",
"sampling_params": ["temperature", "max_tokens"],
"notes": ["No API key required", "Always registered", "No request timeout (local generation can be long-running)", "Image content is flattened to text"]
}
},
"features": {
"default": {
"description": "LLM client — all providers + streaming + embeddings + tool calling",
"dependencies": ["reqwest", "tokio", "serde", "serde_json", "async-trait", "futures", "thiserror", "fastrand", "tracing"]
},
"proxy": {
"description": "Built-in dual-protocol HTTP proxy (OpenAI + Anthropic wire formats)",
"extra_dependencies": ["axum", "tower-http", "bytes"],
"endpoints": ["POST /v1/chat/completions", "POST /v1/messages", "POST /v1/embeddings", "GET /health"],
"auth": "Optional Bearer token (LLMRUST_PROXY_KEY)"
}
},
"embeddings": {
"api": true,
"status": "openai_compatible_and_ollama_implemented",
"provider_implementations": ["openai", "deepseek", "moonshot", "openrouter", "ollama"],
"default_provider_behavior": "unsupported",
"notes": "OpenAI-compatible embeddings transport is implemented. Ollama uses native /api/embed. Actual upstream model support varies by provider/model."
},
"client": {
"model_format": "provider/model",
"routing": {
"description": "LmrsClient parses provider/model prefix to select the registered provider",
"example": "openai/gpt-4o → OpenAIProvider with model 'gpt-4o'"
},
"registry_methods": ["set_openai", "set_anthropic", "set_deepseek", "set_google", "set_ollama", "set_moonshot", "set_openrouter", "set_openai_compatible", "set_custom", "from_env"],
"retry": {
"mechanism": "RetryProvider wraps any Provider with exponential backoff",
"max_retries": "configurable",
"retry_on": ["HTTP 5xx", "429 (rate limit)", "network errors", "stream transient errors"],
"no_retry_on": ["HTTP 4xx (except 429)", "Parse errors", "UnknownProvider errors"]
},
"router": {
"description": "Multi-deployment routing with failover, load balancing, and optional passive cooldown",
"strategies": ["Ordered", "RoundRobin"],
"groups": "Logical group names mapped to ordered deployment lists",
"passive_cooldown": true,
"cooldown_default": "disabled",
"cooldown_policy": "opt-in; failoverable errors mark deployments cooling; all-cooling groups fail open"
}
},
"proxy": {
"description": "Dual-protocol API gateway (feature = 'proxy')",
"openai_endpoint": {
"path": "/v1/chat/completions",
"format": "OpenAI Chat Completions API",
"sse_streaming": true
},
"anthropic_endpoint": {
"path": "/v1/messages",
"format": "Anthropic Messages API",
"sse_streaming": true
},
"embeddings_endpoint": {
"path": "/v1/embeddings",
"protocol": "OpenAI-compatible",
"input": ["string", "string_array"],
"encoding_format": ["float"],
"unsupported": ["base64", "token_array"],
"auth": "same as proxy auth"
},
"health_endpoint": "GET /health",
"auth": {
"type": "Bearer token",
"config": "LLMRUST_PROXY_KEY",
"comparison": "constant-time comparison"
},
"n_policy": "Accepts missing n or n = 1; rejects n = 0 or n > 1 to prevent silent multi-completion billing"
},
"design_principles": [
"default = [] — no optional dependencies unless explicitly requested",
"Native protocols for Anthropic and Gemini (not OpenAI-compatible shims)",
"Stream errors are surfaced, never silently dropped",
"No logging of API keys, prompts, responses, request bodies, tool arguments, image data, embedding input text, embedding vectors, or full URLs",
"All providers implement both chat and stream",
"Provider trait is minimal: required chat() + stream(), optional default embed()"
]
}