kei-registries/providers.toml

# Реестр провайдеров LLM. SSoT для kei-model-router + keisei-marketplace.
#
# Иерархия: transport → provider → model → profile (agent-profiles.toml).
# Транспорт стабилен (direct-api / aws-bedrock / azure-openai / google-vertex / local),
# провайдер стабилен (endpoint, схема аутентификации, биллинг-аккаунт),
# модели за ним меняются — см. models.toml.
#
# Все цены/контекст — в models.toml. Здесь только транспорт + лимиты.
#
# Поле `transport` задаёт способ доступа:
#   - direct-api      — прямой HTTP к провайдеру (Anthropic, OpenAI, xAI, etc.)
#   - aws-bedrock     — через AWS Bedrock (IAM/role, регион)
#   - azure-openai    — через Azure (deployment + Azure key)
#   - google-vertex   — через GCP Vertex AI (service-account JSON + project/region)
#   - local           — локальная LLM (Ollama, MLX, LM Studio)
#   - proxy           — собственный или сторонний прокси (LiteLLM, OpenRouter)
#   - subscription    — OAuth-привязка через подписку (ChatGPT, Claude.ai)
#
# Конвенция: `id` — kebab-case с суффиксом транспорта если не direct-api
# (anthropic, anthropic-bedrock; openai, openai-azure; google, google-vertex).

[[provider]]
id = "anthropic"
display_name = "Anthropic (Direct API)"
transport = "direct-api"
endpoint = "https://api.anthropic.com/v1/messages"
auth_scheme = "x-api-key"          # header name; value from env ANTHROPIC_API_KEY
auth_env = "ANTHROPIC_API_KEY"
api_version_header = "anthropic-version"
api_version_value = "2023-06-01"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 4000
billing_currency = "USD"
notes = "Primary provider for Claude family. RULE 0.20 default for code-implementer-*."

[[provider]]
id = "anthropic-bedrock"
display_name = "Anthropic (AWS Bedrock)"
transport = "aws-bedrock"
endpoint = "https://bedrock-runtime.{region}.amazonaws.com"
auth_scheme = "aws-sigv4"
auth_env = "AWS_ACCESS_KEY_ID,AWS_SECRET_ACCESS_KEY,AWS_REGION"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 0                 # per-region quota; see AWS console
billing_currency = "USD-aws"
notes = "Claude через AWS Bedrock — для enterprise с AWS-биллингом. Регион в AWS_REGION."

[[provider]]
id = "openai"
display_name = "OpenAI (Direct API)"
transport = "direct-api"
endpoint = "https://api.openai.com/v1/chat/completions"
auth_scheme = "bearer"
auth_env = "OPENAI_API_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 5000
billing_currency = "USD"
notes = "GPT family + codex (ChatGPT OAuth, see codex provider for dual-review)."

[[provider]]
id = "openai-azure"
display_name = "OpenAI (Azure)"
transport = "azure-openai"
endpoint = "https://{azure_resource}.openai.azure.com/openai/deployments/{deployment}"
auth_scheme = "api-key"
auth_env = "AZURE_OPENAI_API_KEY,AZURE_OPENAI_ENDPOINT,AZURE_OPENAI_DEPLOYMENT"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 0                 # per-deployment quota; see Azure portal
billing_currency = "USD-azure"
notes = "OpenAI через Azure — для enterprise с Azure-биллингом. Нужны resource+deployment."

[[provider]]
id = "claude-code"
display_name = "Claude Code (subscription — your claude CLI, no API key)"
transport = "subscription"
endpoint = "local:claude-cli"      # invoked via the `claude` CLI (Claude Code), not HTTP
auth_scheme = "oauth-subscription"
auth_env = "_"                     # no env — Claude Code handles auth (claude.ai Max / Pro)
retry_max = 1
retry_backoff_ms = 0
rate_limit_rpm = 60                # Claude subscription quota
billing_currency = "USD-subscription"
notes = "Default for the Claude-Code-native kit. Uses ambient `claude` CLI OAuth; no API key needed."

[[provider]]
id = "codex"
display_name = "OpenAI Codex (ChatGPT OAuth)"
transport = "subscription"
endpoint = "local:codex-cli"       # invoked via `codex` CLI, not HTTP
auth_scheme = "oauth-subscription"
auth_env = "_"                      # no env — CLI handles auth
retry_max = 1
retry_backoff_ms = 0
rate_limit_rpm = 60                # ChatGPT subscription quota
billing_currency = "USD-subscription"
notes = "RULE 0.23 DUAL REVIEW. Separate billing from API providers."

[[provider]]
id = "xai"
display_name = "xAI"
transport = "direct-api"
endpoint = "https://api.x.ai/v1/chat/completions"
auth_scheme = "bearer"
auth_env = "XAI_API_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 1000
billing_currency = "USD"
notes = "Grok family. OpenAI-compatible API."

[[provider]]
id = "deepseek"
display_name = "DeepSeek"
transport = "direct-api"
endpoint = "https://api.deepseek.com/v1/chat/completions"
auth_scheme = "bearer"
auth_env = "DEEPSEEK_API_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 60
billing_currency = "USD"           # invoices in USD even if upstream CNY
notes = "Cheapest tier for batch reasoning. R1 reasoner family."

[[provider]]
id = "google"
display_name = "Google Gemini (Direct API)"
transport = "direct-api"
endpoint = "https://generativelanguage.googleapis.com/v1beta/models"
auth_scheme = "query-key"          # ?key=<...> in URL
auth_env = "GEMINI_API_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 1000
billing_currency = "USD"
notes = "1M-context window for long-doc analysis."

[[provider]]
id = "google-vertex"
display_name = "Google Gemini (Vertex AI)"
transport = "google-vertex"
endpoint = "https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/publishers/google/models"
auth_scheme = "gcp-service-account"
auth_env = "GOOGLE_APPLICATION_CREDENTIALS,GCP_PROJECT_ID,GCP_REGION"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 0                 # per-project quota; see GCP console
billing_currency = "USD-gcp"
notes = "Gemini через Vertex AI — для enterprise с GCP-биллингом. Нужен service-account JSON."

[[provider]]
id = "ollama-local"
display_name = "Ollama (local)"
transport = "local"
endpoint = "http://127.0.0.1:11434/api/chat"
auth_scheme = "none"
auth_env = "_"
retry_max = 1
retry_backoff_ms = 0
rate_limit_rpm = 0                 # local, no remote limit
billing_currency = "USD-zero"      # local compute, opportunity cost only
notes = "Self-hosted. Models loaded via `ollama pull`."

[[provider]]
id = "mlx-local"
display_name = "MLX (Apple silicon local)"
transport = "local"
endpoint = "http://127.0.0.1:8080/v1/chat/completions"
auth_scheme = "none"
auth_env = "_"
retry_max = 1
retry_backoff_ms = 0
rate_limit_rpm = 0
billing_currency = "USD-zero"
notes = "Native Apple-silicon inference. mlx_lm.server."

[[provider]]
id = "lmstudio-local"
display_name = "LM Studio (local)"
transport = "local"
endpoint = "http://127.0.0.1:1234/v1/chat/completions"
auth_scheme = "none"
auth_env = "_"
retry_max = 1
retry_backoff_ms = 0
rate_limit_rpm = 0
billing_currency = "USD-zero"
notes = "Desktop GUI runner. OpenAI-compatible."

[[provider]]
id = "litellm-proxy"
display_name = "LiteLLM proxy (keisei.app)"
transport = "proxy"
endpoint = "https://api.keisei.app/llm/v1/chat/completions"
auth_scheme = "bearer"
auth_env = "KEI_LITELLM_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 1000
billing_currency = "USD-keisei"
notes = "Per-user-keyed proxy. Routes to anthropic/openai/xai/deepseek under one bearer."

[[provider]]
id = "openrouter"
display_name = "OpenRouter (multi-provider proxy)"
transport = "proxy"
endpoint = "https://openrouter.ai/api/v1/chat/completions"
auth_scheme = "bearer"
auth_env = "OPENROUTER_API_KEY"
retry_max = 3
retry_backoff_ms = 500
rate_limit_rpm = 1000
billing_currency = "USD"
notes = "Меняет провайдеров по pricing. Подходит для эксплоративных задач."