Getting StartedQuickstart

Quickstart

Run Agent Flywheel with the verified Python SDK, OpenClaw plugin, or DeepSeek Harness path, with clear boundaries for observation and workflow access.

Start your first Agent Flywheel integration

Choose the integration that matches your runtime and required access boundary.

IntegrationBest fitCapability boundary
Python SDKA Python agent exposed through module:attrServes the agent with text, messages, or dict input
OpenClawA full workflow integrationRequires conversation access enabled as a separate host permission
DeepSeek HarnessObserve-only trackingTracks tools/result outcomes without controlling rollouts or production turns

Setup flow: choose an integration, install its SDK or plugin, configure the required permission or credential, then start the integration using the instructions in its tab.

Use the Python SDK when your agent entry point is available as a module:attr reference. The SDK accepts text, messages, and dict input shapes.

Install the SDK

pip install agent-flywheel

Configure credentials

Set PERCEPTEYE_API_KEY in the environment that starts your agent. Use a key with the pefw prefix.

Serve your agent

Pass your agent's module:attr entry point, identifier, and input shape.

import agent_flywheel as fw

fw.serve("my_package.agent:build", agent_id="my-coding-agent", input_shape="text")

Set input_shape to text, messages, or dict to match the input your agent accepts.