Model Adapters

opentine uses a unified Model protocol that every adapter implements. This means you can swap between Anthropic, OpenAI, Google, Ollama, or any compatible provider by changing a single line of code. Your agent logic, tools, and run tree structure stay exactly the same.

The Model Protocol

Every model adapter implements this protocol. If you're building a custom adapter, this is the interface you need to satisfy.

opentine/models/base.py
1class Model(Protocol):
2    @property
3    def name(self) -> str: ...
4
5    @property
6    def supports_tools(self) -> bool: ...
7
8    @property
9    def supports_thinking(self) -> bool: ...
10
11    async def complete(
12        self,
13        messages: list[dict],
14        tools: list[dict] | None = None,
15        system: str | None = None,
16        temperature: float = 0.0,
17    ) -> dict:
18        # Returns: {"text": str, "tool_calls": list[dict], "cost": float}
19        ...
20
21    async def stream(
22        self,
23        messages: list[dict],
24        tools: list[dict] | None = None,
25        system: str | None = None,
26        temperature: float = 0.0,
27    ) -> AsyncIterator[dict]:
28        # Yields: {"type": "text_delta", "text": str}
29        ...

Properties

  • name — A string identifier for the model (e.g. "anthropic/claude-sonnet-4-20250514", "ollama/llama3.1").
  • supports_tools — Whether the model supports tool calling. All built-in adapters return True.
  • supports_thinking — Whether the model supports extended thinking / chain-of-thought. This varies by model and provider.

Methods

  • complete() — Send messages and get a full response back. Returns a dict with text, tool_calls, and cost.
  • stream() — Send messages and get an async iterator of text deltas. Each chunk is a dict with type and text.

Swapping Models

Because every adapter implements the same protocol, swapping models is a one-line change. Your Agent, tools, and run tree format are completely model-agnostic.

swap_models.py
1from opentine import Agent
2
3# Just change the import and constructor — everything else stays the same
4
5from opentine.models.anthropic import Anthropic
6agent = Agent(model=Anthropic("claude-sonnet-4-20250514"), tools=[...])
7
8# Swap to OpenAI
9from opentine.models.openai import OpenAI
10agent = Agent(model=OpenAI("gpt-4o"), tools=[...])
11
12# Swap to Google
13from opentine.models.google import Google
14agent = Agent(model=Google("gemini-2.0-flash"), tools=[...])
15
16# Swap to a local model via Ollama
17from opentine.models.ollama import Ollama
18agent = Agent(model=Ollama("llama3.1"), tools=[...])

Using Complete

The complete() method is the primary interface for getting a full response from a model. It returns the response text, any tool calls, and the cost of the request.

complete_example.py
1from opentine.models.anthropic import Anthropic
2
3model = Anthropic("claude-sonnet-4-20250514")
4
5# Use complete() for a single response
6result = await model.complete(
7    messages=[{"role": "user", "content": "Explain quantum entanglement"}],
8    system="You are a physics teacher.",
9    temperature=0.0,
10)
11
12print(result["text"])           # The model's response text
13print(result["tool_calls"])     # Any tool calls (empty list if none)
14print(result["cost"])           # Cost in USD for this call

Using Stream

The stream() method returns an async iterator that yields text deltas as the model generates them. This is useful for real-time output or long responses.

stream_example.py
1from opentine.models.openai import OpenAI
2
3model = OpenAI("gpt-4o")
4
5# Use stream() for incremental output
6async for chunk in model.stream(
7    messages=[{"role": "user", "content": "Write a haiku about Python"}],
8):
9    print(chunk["text"], end="", flush=True)

Adapter Comparison

AdapterDefault ModelToolsThinkingEnv Variable
Anthropicclaude-sonnet-4-20250514YesYes (Opus, Sonnet)ANTHROPIC_API_KEY
OpenAIgpt-4oYesYes (o-series)OPENAI_API_KEY
Googlegemini-2.0-flashYesYes (thinking models)GOOGLE_API_KEY
Ollamallama3.1YesNoOLLAMA_HOST
LiteLLM———Coming soon

Next Steps