Providers
Clew supports 27 AI providers with a unified adapter layer. Set your API keys as environment variables and switch between providers at runtime with /model.
Provider configuration is declared in src/services/ai/providers.json. The ProviderManager resolves API keys, selects models, and manages provider config. Non-Anthropic providers are normalized via the AnthropicAdapter or GoogleAdapter.
All Providers
| Provider | Env Key | Default Model | Notes |
|---|---|---|---|
| Anthropic | ANTHROPIC_API_KEY | claude-sonnet-4-7 | Native Claude support with full tool calling and reasoning. Context: 1M tokens. |
| OpenAI | OPENAI_API_KEY | gpt-5.5 | GPT-5.5, GPT-5.5 Pro, GPT-5.4 Mini. Context: 1M+ tokens. |
GOOGLE_API_KEY | gemini-3.1-flash | Gemini 3.1 Flash, Pro. Context: up to 2M tokens. Uses GoogleAdapter. | |
| GitHub Copilot | COPILOT_GITHUB_TOKEN | gpt-5.5 | Access via GitHub token. Includes Claude and GPT models. Context: 1M tokens. |
| OpenRouter | OPENROUTER_API_KEY | — | 100+ models via routed API. OpenAI-compatible. |
| DeepSeek | DEEPSEEK_API_KEY | deepseek-v4-pro | DeepSeek V4 Pro, V4 Flash. Context: 1M tokens. MoE architecture. |
| OpenCode | OPENCODE_API_KEY | claude-opus-4-7 | OpenCode AI gateway. OpenAI-compatible. Includes Claude, GPT, Gemini, Kimi, GLM, MiniMax. |
| OpenCode Go | OPENCODE_GO_API_KEY | minimax-m2.7 | Low cost subscription tier for open coding models. |
| Cline API | CLINE_API_KEY | anthropic/claude-opus-4-7 | OpenAI-compatible. Model IDs format: provider/model-name. |
| Groq | GROQ_API_KEY | — | Llama 4 70B, 8B. Fast inference. Context: 128K tokens. |
| xAI | XAI_API_KEY | grok-4-20 | Grok 4.20, Grok 4. Context: 128K tokens. Vision and tool calling. |
| Mistral | MISTRAL_API_KEY | mistral-small-latest | Mistral Large 4, Small 4, Ministral 3 14B. Tool calling and vision support. |
| KiloCode | KILOCODE_API_KEY | kilo-auto/free | Gateway with 500+ models. Includes free tier. |
| Ollama (Local) | OLLAMA_API_KEY | llama3.3 | Local models via Ollama server at http://localhost:11434/v1. Offline-capable. |
| Together AI | TOGETHER_API_KEY | — | Llama 4 models. OpenAI-compatible. |
| Fireworks AI | FIREWORKS_API_KEY | — | Llama 4 70B. Fast inference. |
| NVIDIA NIM | NVIDIA_API_KEY | — | NVIDIA AI Foundation Models. DeepSeek, Llama 3.1, Nemotron. |
| Cohere | COHERE_API_KEY | command-r-plus | Command R+ model. 128K context. No vision. |
| Deep Infra | DEEPINFRA_API_KEY | — | OpenAI-compatible gateway. Llama 4 Scout with 327K context. |
| Perplexity | PERPLEXITY_API_KEY | — | Sonar Huge Online. No tool calling support. |
| Cerebras | CEREBRAS_API_KEY | — | Wafer-Scale Engine. Extremely fast inference. Qwen 3 235B. |
| SiliconFlow | SILICONFLOW_API_KEY | — | China-based AI platform. DeepSeek V4 Pro and Flash. |
| Moonshot AI (Kimi) | MOONSHOT_API_KEY | kimi-latest | Kimi models. 128K context. Vision and tool calling. |
| Zhipu AI (GLM) | ZHIPU_API_KEY | glm-4 | GLM-4 model. 128K context. |
| Hugging Face | HUGGINGFACE_API_KEY | — | Inference API. Llama 3.1 70B. No tool calling. |
| Poe | POE_API_KEY | — | OpenAI-compatible gateway. Claude 3.7 Sonnet available. |
| DigitalOcean | DIGITALOCEAN_API_KEY | — | Serverless Inference. Llama 3.1 70B. |
Provider Selection
ProviderManager resolves the active provider by:
- Scanning environment variables for configured API keys
- Loading provider metadata from
providers.json - Resolving model names to provider-specific identifiers
- Applying the adapter layer for non-Anthropic providers
- Handling config migration and backward compatibility
Adapter Layer
Non-Anthropic providers pass through adapters that normalize:
- Content blocks — text, tool_use, tool_result, thinking block conversion (
contentBlockUtils.ts) - Tool calls — parsing across provider formats (
toolCallParser.ts) - Streaming chunks — per-chunk normalization for real-time display
- Errors — normalization with retry/rate-limit handling (
errorNormalizer.ts) - Usage — token counting and cost estimation (
usageNormalizer.ts)
Runtime Switching
/model # Interactive model picker
/model list # List all available models across providers
/model claude-sonnet-4-7 # Switch to a specific model
/provider-select # Show or change active providerProvider Capabilities
Each provider declares capabilities in providers.json:
- chat — Chat completion endpoint
- streaming — full, partial, or none
- toolCalling — native, adapted, or none
- vision — Image input support
- jsonSchema — Structured output via JSON schema
- reasoningEffort — Configurable reasoning depth
- contextLength — Maximum context window size
Local / Offline Development
# Ollama — fully offline
export OLLAMA_API_KEY=ollama # (any non-empty value)
export OLLAMA_HOST=http://localhost:11434
clew
> /model ollama/llama4:70b
# Debug specific provider
DEBUG=provider:anthropic clew