Providers

CrabTalk supports multiple LLM providers through a unified Model trait. All providers are API-based — configure an endpoint, point the daemon at it, and go. Use crabtalk auth for interactive setup, or edit crab.toml directly.

API standards

Each provider uses a wire format selected by the standard field:

If standard is omitted, CrabTalk defaults to openai_compat. If the base_url contains "anthropic", the Anthropic standard is auto-detected.

Provider configuration

Each provider is a [provider.<name>] section in crab.toml. A provider owns one or more models:

[provider.deepseek]
models = ["deepseek-chat", "deepseek-thinking"]
api_key = "sk-..."

Model names must be unique across all providers.

Selecting the active model

Set the default model in [system.crab]:

[system.crab]
model = "deepseek-chat"

Override per agent in [agents.*]:

[agents.researcher]
model = "claude-opus-4-6"

Supported providers

OpenAI

[provider.openai]
models = ["gpt-4o", "gpt-4o-mini"]
api_key = "sk-..."

Uses the OpenAI chat completions API (the default standard). Supports GPT-4o, GPT-4o-mini, o-series, and all OpenAI models.

Anthropic Claude

[provider.anthropic]
models = ["claude-sonnet-4-20250514", "claude-opus-4-6"]
api_key = "sk-ant-..."
standard = "anthropic"

Uses the Anthropic Messages API. Set standard = "anthropic" explicitly, or it will be auto-detected if your base_url contains "anthropic".

DeepSeek

[provider.deepseek]
models = ["deepseek-chat", "deepseek-thinking"]
api_key = "sk-..."

OpenAI-compatible API. Supports deepseek-chat and deepseek-thinking.

Google Gemini

[provider.google]
models = ["gemini-2.5-pro"]
api_key = "..."
standard = "google"

Grok

[provider.grok]
models = ["grok-3"]
api_key = "..."
base_url = "https://api.x.ai/v1"

OpenAI-compatible API. Requires explicit base_url.

Qwen

[provider.qwen]
models = ["qwen-plus"]
api_key = "..."
base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1"

OpenAI-compatible via DashScope.

Kimi

[provider.kimi]
models = ["kimi-latest"]
api_key = "..."
base_url = "https://api.moonshot.cn/v1"

OpenAI-compatible API by Moonshot AI.

Ollama

[provider.ollama]
models = ["llama3.1"]
standard = "ollama"
base_url = "http://localhost:11434/v1"

No API key needed.

Azure OpenAI

[provider.azure]
models = ["gpt-4o"]
api_key = "..."
standard = "azure"
api_version = "2024-02-01"
base_url = "https://your-resource.openai.azure.com"

AWS Bedrock

[provider.bedrock]
models = ["anthropic.claude-3-5-sonnet"]
standard = "bedrock"
region = "us-east-1"
access_key = "..."
secret_key = "..."

Custom OpenAI-compatible endpoints

Any OpenAI-compatible API works with a base_url:

[provider.my-provider]
models = ["my-model"]
api_key = "..."
base_url = "https://my-endpoint.com/v1"

Provider manager

The ProviderManager holds all configured providers and routes requests by model name. It supports hot-reload — update the config and the active provider changes without restarting the daemon.

What's next

• Configuration — full config setup
• Commands — telegram, search, and custom commands
• Auth — manage providers interactively