Integrations
Choose the Agent Flywheel integration that fits your agent host, from SDK-based capture and local hooks to approval-gated training workflows.
Choose by integration surface
| Surface | Recommended when | Boundary |
|---|---|---|
| Python SDK | Your agent runs in Python and needs runtime instrumentation, training rollouts, or production capture. | Runs inside the agent process. |
| Node plugin | Your host is OpenClaw or DeepSeek Harness on Node >=22.22.3. | Capabilities depend on the host. |
| goose or Gemini CLI hooks | You need local, observe-only tool activity recording. | Writes trajectory data locally. |
| Canonical Agent Skill | An operator needs to manage Mission Control training from a compatible Agent Skills client. | Runs approval-gated, server-side training. |
| Claude Code marketplace plugin | You use Claude Code and want the canonical Agent Skill distributed through its marketplace. | Distribution wrapper, not a separate SDK. |
Need in-process runtime? → Python SDK
Need a host plugin? → OpenClaw full workflow or DeepSeek observe-only
Need local-only evidence? → goose or Gemini CLI hooks
Need operator training controls? → Agent Skill, with the Claude marketplace plugin for Claude Code
Host capabilities differ. Review the relevant section before choosing an integration if you need production capture, training rollouts, or prompt and model application.
Python SDK
Use the agent-flywheel Python SDK when instrumentation belongs inside the agent process. It supports LangChain, LangGraph through LangChain callbacks, Google ADK, arbitrary CLI agents, and custom adapters.
| Method | Use it for | Boundary |
|---|---|---|
serve() | Run training rollouts. | Training execution inside the agent integration. |
attach() | Capture production activity. | Production activity capture inside the agent process. |
record_tool_call() | Record an explicit tool outcome. | One explicit tool-call result. |
Node plugin
Use @percepteye/agent-flywheel with OpenClaw or DeepSeek Harness on Node >=22.22.3.
OpenClaw supports tool outcome recording, training rollouts, production capture, and applying approved prompts and models. Enable conversation access explicitly before the plugin can observe conversation data.
openclaw plugins install @percepteye/agent-flywheel
DeepSeek Harness observes tools/result outcomes only. It does not provide training rollouts, production-turn capture, or prompt and model application.
Install the plugin for a named profile:
dsh plugin --profile <name> add @percepteye/agent-flywheel
goose and Gemini CLI hooks
Use these hooks for local, observe-only tool activity recording. They use the Python standard library only and write trajectory-compatible tool_calls.jsonl data.
Install the hook with:
goose plugin install https://github.com/percepteye-ai/agent-flywheel-hooks
Install the hook with:
gemini extensions install https://github.com/percepteye-ai/agent-flywheel-hooks
Set PERCEPTEYE_TRAJECTORY_DIR to define where the hook writes trajectory data. Replace /path/to/trajectory-directory with the directory you want to use. The example does not prescribe a specific path.
PERCEPTEYE_TRAJECTORY_DIR=/path/to/trajectory-directory
Without PERCEPTEYE_TRAJECTORY_DIR, the recorder remains inert.
These hooks record local tool outcomes only. They do not upload data to Mission Control, claim training rollouts, or apply a model or prompt.
Tool outcome mapping
| Client | Event | Recorded outcome |
|---|---|---|
| goose | PostToolUse | ok |
| goose | PostToolUseFailure | failed |
| Gemini CLI | AfterTool with an error | failed |
| Gemini CLI | AfterTool without an error | unknown |
Reporting excludes unknown outcomes from the denominator. An unclassified Gemini CLI result is not counted as either a success or a failure.
Agent Skills integrations
Canonical Agent Skill
Use percepteye-agent-training for the operator-side Mission Control workflow in compatible Agent Skills clients, including:
- Codex
- Cursor
- Gemini CLI
- GitHub Copilot and VS Code
- Claude
- Other compatible Agent Skills clients
The skill is approval-gated. It uses OTEL GenAI traces or data produced by the SDK, launches and monitors server-side training, and handles approval, result, and promotion actions.
The skill does not train models locally.
Claude Code marketplace plugin
The Claude Code marketplace plugin is a vendored distribution wrapper for the canonical percepteye-agent-training skill. It is not a separate SDK or training workflow.
Install the marketplace and plugin with:
claude plugin marketplace add percepteye-ai/percepteye-claude-plugin
claude plugin install percepteye@percepteye
Invoke the skill in Claude Code with:
/percepteye:percepteye-agent-training
The plugin requires Python >=3.9, HTTPS access to Mission Control, and a Mission Control account. Configure credentials using the Authentication documentation.