Skip to content

Latest commit

 

History

History
114 lines (80 loc) · 4.25 KB

providers.md

File metadata and controls

114 lines (80 loc) · 4.25 KB

Providers

Source of truth:

  • crates/config/src/lib.rs (BUILT_IN_PROVIDERS)
  • crates/provider/src/lib.rs
  • provider implementations under crates/provider/src/*

Provider Model

The runtime uses a provider abstraction:

  • trait: Provider
  • creation: create_provider / create_provider_async
  • capability metadata: ModelInfo, ModelTier, ThinkingSupport

Each provider has:

  • provider id
  • auth path (API key, OAuth, local token)
  • API style (openai, anthropic, gemini, cursor, copilot, claude-sdk, codex)
  • model inventory (hardcoded + optional remote refresh)

Built-in Providers

Provider id Name API style Env var OAuth support Default base URL
openai OpenAI openai OPENAI_API_KEY yes https://api.openai.com/v1
anthropic Anthropic anthropic ANTHROPIC_API_KEY yes https://api.anthropic.com/v1
gemini Google Gemini gemini GEMINI_API_KEY yes https://generativelanguage.googleapis.com/v1beta
cursor Cursor cursor CURSOR_API_KEY yes https://api2.cursor.sh
github-copilot GitHub Copilot copilot GITHUB_COPILOT_TOKEN yes https://api.githubcopilot.com
openrouter OpenRouter openai OPENROUTER_API_KEY no https://openrouter.ai/api/v1
claude-sdk Claude Agent SDK claude-sdk ANTHROPIC_API_KEY no empty (resolved at runtime)
codex OpenAI Codex CLI codex CODEX_API_KEY yes empty (resolved at runtime)
groq Groq openai GROQ_API_KEY no https://api.groq.com/openai/v1
together Together AI openai TOGETHER_API_KEY no https://api.together.xyz/v1
deepseek DeepSeek openai DEEPSEEK_API_KEY no https://api.deepseek.com/v1
ollama Ollama (local) openai OLLAMA_API_KEY no http://localhost:11434/v1
kimi Kimi (Moonshot) openai MOONSHOT_API_KEY no https://api.moonshot.ai/v1
kimi-coding Kimi Coding Plan anthropic KIMI_CODING_API_KEY no https://api.kimi.com/coding
minimax MiniMax openai MINIMAX_API_KEY no https://api.minimax.io/v1
minimax-coding MiniMax Coding Plan anthropic MINIMAX_CODING_API_KEY no https://api.minimax.io/anthropic
glm GLM (Z.ai) openai ZHIPU_API_KEY no https://api.z.ai/api/paas/v4
glm-coding GLM Coding Plan openai ZHIPU_CODING_API_KEY no https://api.z.ai/api/coding/paas/v4

Runtime Provider Resolution

Provider selection path:

  1. CLI --provider if set
  2. config.provider.default otherwise

Model selection path:

  1. CLI --model if set
  2. provider entry model if set
  3. provider first supported model fallback

Special Cases

claude-sdk

  • resolves credential for claude-sdk
  • if unavailable, falls back to anthropic credential
  • may use anthropic base URL fallback

codex

  • resolves credential for codex
  • if unavailable, falls back to openai credential
  • may use openai base URL fallback

cursor

  • credential is parsed into token and machine id via cursor OAuth token parser

github-copilot

  • loads stored token from auth store under github-copilot
  • requires refresh token path for runtime provider instantiation

Model Registry and Tiering

ModelRegistry tracks hardcoded model lists by provider and supports:

  • models_for(provider)
  • find(provider, model_id)
  • find_any("provider/model") or bare model id
  • list providers and all models

Remote model refresh flow:

  • refresh_provider_models(provider_id, cache)
  • attempts provider model listing API
  • merges fetched models with hardcoded models
  • falls back to hardcoded list if fetch fails

Routing Integration

When routing is enabled (agent.routing.enabled=true):

  • prompt is classified into tier (low|medium|high)
  • provider model_for_tier selects best matching model
  • fallback is provider first model

See docs/routing.md.

Practical Recommendations

  • Use explicit provider/model notation in high-control workflows.
  • Pin model under [provider.<id>] for reproducibility.
  • Keep API keys out of repository files; prefer env vars or OAuth.
  • For local/offline style workflows, configure ollama with local runtime.