Providers¶

Providers are the AI backends that power Missy's responses. The provider system is built around a registry pattern with a common interface, so switching between Claude, GPT, Ollama, or any OpenAI-compatible service requires only a config change.

Supported providers¶

Provider	Module	API key required	Local	Tool calling
Anthropic	`AnthropicProvider`	Yes	No	Yes
OpenAI	`OpenAIProvider`	Yes	No	Yes
OpenAI Codex	`CodexProvider`	OAuth token	No	Yes
Ollama	`OllamaProvider`	No	Yes	Yes

How it works¶

ProviderRegistry¶

The ProviderRegistry is the single source of truth for active providers. On startup, Missy reads the providers: section of your config and instantiates each enabled provider:

providers:
  anthropic:
    name: anthropic
    model: "claude-sonnet-4-6"
    api_key: null                  # Falls back to ANTHROPIC_API_KEY env var
    enabled: true
  ollama:
    name: ollama
    model: "llama3.2"
    base_url: "http://localhost:11434"
    enabled: true

Check what is registered and available:

missy providers

BaseProvider interface¶

All providers implement the same interface:

Method	Description
`complete(messages)`	Send conversation turns, get a response
`complete_with_tools(messages, tools, system)`	Completion with tool-calling support
`stream(messages, system)`	Stream partial response tokens
`is_available()`	Check if the provider is ready (credentials present, service reachable)
`get_tool_schema(tools)`	Convert tools to provider-native schema format

The common data types:

Message(role="user", content="Hello")

CompletionResponse(
    content="Hi there!",
    model="claude-sonnet-4-6",
    provider="anthropic",
    usage={"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
    tool_calls=[],            # List[ToolCall] when model requests tool use
    finish_reason="stop",     # "stop" | "tool_calls" | "length"
)

ToolCall(id="call_123", name="calculator", arguments={"expression": "2+2"})
ToolResult(tool_call_id="call_123", name="calculator", content="4")

ModelRouter¶

The ModelRouter automatically selects between fast, primary, and premium model tiers based on prompt complexity:

Tier	When used	Example model
`fast`	Short queries, simple lookups	`claude-haiku-4-5`
`primary`	General purpose (default)	`claude-sonnet-4-6`
`premium`	Complex tasks, long context, many tools	`claude-opus-4-6`

Configure tier models per provider:

providers:
  anthropic:
    name: anthropic
    model: "claude-sonnet-4-6"
    fast_model: "claude-haiku-4-5"
    premium_model: "claude-opus-4-6"

API key rotation¶

Providers support multiple API keys for load distribution and rate limit management:

providers:
  anthropic:
    name: anthropic
    model: "claude-sonnet-4-6"
    api_keys:
      - "sk-ant-key1..."
      - "sk-ant-key2..."
      - "sk-ant-key3..."

Keys rotate round-robin when the registry's rotate_key() method is called (typically after rate limit errors).

Network policy¶

All provider HTTP traffic passes through the PolicyHTTPClient, which enforces network policy. Provider hosts are automatically added to provider_allowed_hosts from their base_url configuration, so you generally do not need to add them manually.

If using default_deny: true and providers are not auto-detected, add them explicitly:

network:
  provider_allowed_hosts:
    - "api.anthropic.com"
    - "api.openai.com"