Python SDK

Courier SDK

Develop robust and reliable programmatic agents with our simple-to-use agent SDK that manages agent loops.

Write once. Run either place.

You should not rewrite an agent because the model moved from a token API to a Mac on your desk. Courier SDK is the same Model, Agent, and Operator — two runtimes.

from courier_os import Model, Agent, Operator

Same Python. Two runtimes.

API mode

Point at Courier Cloud, OpenAI, or any OpenAI-compatible base URL. No binary, no license — an API key.

pip install courier-os

Local mode

Load an MLX model on Apple Silicon you own. Same agent code. Unlimited usage, air-gapped.

Contact for License

What ships in the package

Python agents you can actually ship: tools, a tool loop, a session that can grow mid-run, and a one-shot relay.

Model, Agent, Operator

Agent is one model turn — you handle tool_calls. Operator runs the loop until the model finishes.

@tools.tool

Typed tools you own. Decorate a function; the SDK ships the OpenAI tool schema.

Session + mid-loop register

Pass a Session and tools=session.tools, then event.register_tool from an intercept. New tools hit the next Operator iteration.

relay()

One-shot HTTP path: model, messages, tools, api_key, base_url. Same surface as Agent without the object graph.

Multimodal

OpenAI-shaped image and audio content parts on both paths. Local omni models extract media before inference.

courier-os[local]

Optional courier-os-binary on Apple Silicon. Same Model, Agent, and Operator — air-gapped MLX.

# pip install courier-os
# or contact for local inference license  # Apple Silicon
from courier_os import Model, Agent, Operator, tools

@tools.tool
def get_weather(city: str) -> dict:
    return {"city": city, "temp_f": 72}

model = Model("gpt-4o", api_key="…", base_url="https://api.openai.com/v1")
# model = Model("mlx-community/…", mode="local")

agent = Agent(model, tools=[get_weather])
out = Operator(agent).run([{"role": "user", "content": "Weather in Denver?"}])
print(out.content)

Full README, dual-mode table, Session intercepts, and multimodal examples on PyPI.