The SDK to build AI agents as orchestratable, event-driven microservices.

Calfkit lets you compose agents as decoupled microservices–agents, tools, workflows–that communicate asynchronously. Add agents to teams without hardcoding any orchestration logic, scale each component independently, and stream agent outputs to any downstream listener.

pip install calfkit

Building agents like traditional web applications—tight coupling and synchronous API calls—creates the same scalability problems that plagued early microservices:

• Tight coupling: Changing one tool or agent breaks dependent agents and tools
• Scaling bottlenecks: All agents and tools live on one runtime, so everything must scale together
• Siloed: Agent communication models are difficult to wire into existing upstream and downstream systems
• Non-streaming: Agents do not naturally follow a livestreaming pattern, making data stream consumption difficult to manage

Calfkit is a Python SDK that builds event-stream agents out-the-box. You get the benefits of an asynchronous, distributed system without managing the infrastructure yourself.

• Distributed to the core: Agents aren't monoliths that just sit on top of the transportation layer. Agents are decomposed into independent services — the agent itself is a deeply distributed system.

• Independent scaling: Each service can scale on its own based on demand.

• Livestream agents by default: Agents already listen on event streams, so consuming data streams — realtime market feeds, IoT sensors, user activity event streams — is the native pattern, not a bolted-on integration.

• Compose agents without coupling: Compose multi-agent teams and workflows by deploying agents on communication channels that are already tapped into the messaging stream. No extra wiring, and no editing existing code — agents don't even need to know about each other.

• Universal data flow: Agents plug into any stream — integrate and consume from any upstream data sources and publish to downstream systems like CRMs, warehouses, or even other agents.

• Python 3.10 or later
• Docker installed and running (for testing with a local Calfkit broker)
• LLM Provider API key

pip install calfkit

git clone https://github.com/calf-ai/calfkit-broker && cd calfkit-broker && make dev-up

Define and deploy a tool as an independent service. Tools are not owned by or coupled to any specific agent—once deployed, any agent in your system can discover and invoke the tool. Deploy once, use everywhere.

# weather_tool.py
import asyncio
from calfkit.nodes import agent_tool
from calfkit.client import Client
from calfkit.worker import Worker

# Define a tool — the @agent_tool decorator turns any function into a deployable tool node
@agent_tool
def get_weather(location: str) -> str:
    """Get the current weather at a location"""
    return f"It's sunny in {location}"

async def main():
    client = Client.connect("localhost:9092")  # Connect to Kafka broker
    worker = Worker(client, nodes=[get_weather])  # Initialize a worker with the tool node
    await worker.run()  # (Blocking call) Deploy the service to start serving traffic

if __name__ == "__main__":
    asyncio.run(main())

Run the file to deploy the tool service:

python weather_tool.py

Deploy the agent as its own service. The Agent handles LLM chat, tool orchestration, and conversation management in a single node. Import the tool definition to register it with the agent—the tool definition is reusable and does not couple the agent to the tool's deployment.

# agent_service.py
import asyncio
from calfkit.nodes import Agent
from calfkit.providers import OpenAIResponsesModelClient
from calfkit.client import Client
from calfkit.worker import Worker
from weather_tool import get_weather  # Import the tool definition (reusable)

agent = Agent(
    "weather_agent",
    system_prompt="You are a helpful assistant.",
    subscribe_topics="weather_agent.input",
    model_client=OpenAIResponsesModelClient(model_name="gpt-5.4-nano"),
    tools=[get_weather],  # Register tool definitions with the agent
)

async def main():
    client = Client.connect("localhost:9092")  # Connect to Kafka broker
    worker = Worker(client, nodes=[agent])  # Initialize a worker with the agent node
    await worker.run()  # (Blocking call) Deploy the service to start serving traffic

if __name__ == "__main__":
    asyncio.run(main())

Set your OpenAI API key:

export OPENAI_API_KEY=sk-...

Run the file to deploy the agent service:

python agent_service.py

Send a request and receive the response. The Client handles broker communication and request correlation automatically.

# invoke.py
import asyncio
from calfkit.client import Client

async def main():
    client = Client.connect("localhost:9092")  # Connect to Kafka broker

    # Send a request and await the response
    result = await client.execute_node(
        "What's the weather in Tokyo?",
        "agent.input",  # The topic the agent subscribes to
    )
    print(f"Assistant: {result.output}")

if __name__ == "__main__":
    asyncio.run(main())

Run the file to invoke the agent:

python invoke.py

Agents can be deployed with a final_output_type to enforce structured output from the LLM. The output is type-safe and deserialized automatically on the client side.

from dataclasses import dataclass
from calfkit.nodes import Agent
from calfkit.providers import OpenAIResponsesModelClient

@dataclass
class WeatherReport:
    location: str
    summary: str

agent = Agent(
    "weather_agent",
    system_prompt="You are a helpful assistant.",
    subscribe_topics="weather_agent.input",
    model_client=OpenAIResponsesModelClient(model_name="gpt-5.4-nano"),
    final_output_type=WeatherReport,  # Enforce structured output
)

When invoking, pass the matching output_type to deserialize the response:

result = await client.execute_node(
    "What's the weather in Tokyo?",
    "weather_agent.input",
    output_type=WeatherReport,
)
print(result.output.location)  # "Tokyo"
print(result.output.summary)   # "It's sunny in Tokyo"

The Client supports multi-turn conversations, runtime dependency injection, and temporary instruction overrides—all without redeploying the agent.

Multi-turn conversations — pass the message history from a previous result to maintain context:

result = await client.execute_node("What's the weather in Tokyo?", "agent.input")

# Continue the conversation with full context
result = await client.execute_node(
    "How about in Osaka?",
    "agent.input",
    message_history=result.message_history,
)

Runtime dependency injection — pass runtime data to tools via the deps parameter:

result = await client.execute_node(
    "What's my phone number?",
    "agent.input",
    deps={"user_id": "usr_123"},  # Available to tools via ctx.deps.provided_deps
)

Temporary instructions — temporarily add system-level instructions scoped per request:

result = await client.execute_node(
    "What's the weather in Tokyo?",
    "agent.input",
    temp_instructions="Always respond in Japanese.",
)

When multiple agents share an input topic (each with its own consumer group), every agent receives every message published to that topic. A gate stack lets a node decide whether to handle an inbound event before run() runs — avoiding wasted LLM tokens on messages addressed elsewhere.

Gates are predicates: Callable[[SessionRunContext], bool | Awaitable[bool]]. They stack with AND semantics in registration order and short-circuit on the first False, exception, or non-bool return. When any gate rejects, run() is skipped and the envelope is returned unchanged — the Kafka offset still commits.

Constructor form — good for shared, cross-cutting predicates passed in as values:

def is_scheduler_target(ctx) -> bool:
    discord = ctx.deps.provided_deps.get("discord", {})
    return discord.get("slash_target") == "scheduler"

scheduler = Agent(
    "scheduler",
    subscribe_topics="discord.thread.123",
    model_client=OpenAIResponsesModelClient(model_name="gpt-5.4-nano"),
    gates=[is_scheduler_target],
)

Decorator form — good for node-specific gates defined inline:

scheduler = Agent("scheduler", subscribe_topics="discord.thread.123", model_client=...)

@scheduler.gate
def is_scheduler_target(ctx) -> bool:
    discord = ctx.deps.provided_deps.get("discord", {})
    return discord.get("slash_target") == "scheduler"

Constructor and decorator forms can be combined; constructor gates run first.

Idempotency requirement: Kafka may redeliver an event before its offset commits, so gates may run more than once for the same logical message. Keep gate functions deterministic and side-effect-free.

Failure behavior: If a gate raises or returns a non-bool, the framework logs the failure and rejects the message (fail-safe). Place cheap fast-reject gates first to maximize short-circuit efficiency.

For tool-node gating, pass gates=[...] to ToolNodeDef.create_tool_node(...) directly; the @agent_tool decorator doesn't expose gates= because tool topics are typically 1:1.

Full documentation is coming soon. In the meantime, this README serves as the primary reference for getting started with Calfkit.

If you found this project interesting or useful, please consider:

• ⭐ Starring the repository — it helps others discover it!
• 🐛 Reporting issues
• 🔀 Submitting PRs

This project is licensed under the Apache License 2.0. See the LICENSE file for details.