Funky Python SDK

The Python client for building and running agents with Funky. It provides matching synchronous and asynchronous APIs for agents, environments, sessions, and session events.

The SDK is currently an alpha. It covers Funky's /v1 data-plane API and intentionally does not expose organization, project, membership, API-key, or other /console/v1 administration APIs.

Installation

pip install funky-sdk

Funky supports Python 3.10 and newer.

Quick start

Set an API key provisioned through Funky:

export FUNKY_API_KEY=fk_...

Then create reusable agent and environment configurations, and start a session:

from funky import Funky

with Funky() as client:
    agent = client.agents.create(
        name="Repository investigator",
        system_prompt="You are a careful coding agent. Verify claims with tools.",
        model={"provider": "anthropic", "model": "claude-sonnet-5"},
        tool_policy={"max_iterations": 20},
    )

    environment = client.environments.create(
        name="github-access",
        network={
            "type": "limited",
            "allowed_hosts": ["github.com", "*.githubusercontent.com"],
        },
    )

    session = client.sessions.create(
        agent=agent.id,
        environment_id=environment.id,
        title="Investigate issue 42",
    )
    client.sessions.wait_until_ready(session.id, timeout=180)

    result = client.sessions.run_turn(
        session.id,
        content="Inspect the repository and identify the root cause.",
    )
    print(result.output_text)

Create agents and environments during application setup and retain their IDs. A new session is appropriate for each independent run or durable conversation.

Async client

AsyncFunky exposes the same resources and method names. Network operations are awaitable and pagination uses async for. run_turn() manages the event stream internally, so the common request-response path only needs the client context:

from funky import AsyncFunky


async def ask_agent(session_id: str) -> None:
    async with AsyncFunky() as client:
        result = await client.sessions.run_turn(
            session_id,
            content="Summarize the repository.",
        )
        print(result.output_text)

run_turn() returns a RunTurnResult containing output_text, the message submission, all events observed during the turn, and the terminal event. It raises TurnFailedError when the agent reports a failed turn. Use stream_events() directly when an application needs to process assistant messages or tool results in real time.

Resource APIs

The synchronous client exposes:

client.agents.create(...)
client.agents.list(...)
client.agents.iter(...)
client.agents.retrieve(agent_id)
client.agents.update(agent_id, ...)
client.agents.archive(agent_id)
client.agents.list_versions(agent_id, ...)
client.agents.iter_versions(agent_id, ...)
client.agents.retrieve_version(agent_id, version)

client.environments.create(...)
client.environments.list(...)
client.environments.iter(...)
client.environments.retrieve(environment_id)
client.environments.update(environment_id, ...)
client.environments.archive(environment_id)
client.environments.delete(environment_id)

client.sessions.create(...)
client.sessions.list(...)
client.sessions.iter(...)
client.sessions.retrieve(session_id)
client.sessions.archive(session_id)
client.sessions.send_message(session_id, content=...)
client.sessions.run_turn(session_id, content=...)
client.sessions.list_events(session_id, ...)
client.sessions.iter_events(session_id, ...)
client.sessions.stream_events(session_id, ...)
client.sessions.wait_until_ready(session_id, ...)
client.sessions.wait_for_turn(session_id, ...)

All API responses are dataclass models. Configuration inputs accept dictionaries; agent models and references also have typed ModelConfig, RuntimeConfig, and AgentReference forms.

Unknown future session-event and content-block variants are returned as UnknownSessionEvent and UnknownContentBlock rather than causing parsing failures. Their original fields are available through each model's raw attribute.

Pagination

Use list() when page metadata matters, or the auto-paginating iterator:

for agent in client.agents.iter(include_archived=True):
    print(agent.id, agent.name)

for event in client.sessions.iter_events(session_id, after_seq=0):
    print(event.seq, event.type)

Client configuration

import httpx
from funky import Funky

http_client = httpx.Client(transport=my_transport)
client = Funky(
    api_key="fk_...",  # defaults from FUNKY_API_KEY
    base_url="https://api.funky.dev",
    timeout=30,
    max_retries=2,
    user_agent="my-service/1.0",
    http_client=http_client,  # useful for custom transports and tests
)

A supplied HTTP client remains owned by the caller and is not closed by Funky. The SDK never includes the API key in its representation or exceptions.