Wake docs

Wake is security for AI agents: it records what every agent does, traces what led to each action, stops agents when it matters, and proves what happened. These docs cover its open source recorder, the part that runs in your stack: the SDKs inside your agents, the proxy that records MCP and A2A calls, the collector that signs what they send, and the self-hosted signed log. It's all in usewake/wake, under Apache 2.0.

Start here

• Quickstart: run the reference fleet on your machine. Four agents on three frameworks talk over A2A and MCP and are recorded end to end. It needs Docker and Python 3, and the agents run on scripted models, so there are no API keys to set.
• Run the recorder: one Compose file runs the collector, the proxy and agentgateway in one place, such as a staging environment.

Add it to your agents

• Python SDK and TypeScript SDK: record every model call, tool call, handoff and memory read or write, for the Claude Agent SDK, the OpenAI Agents SDK, LangChain and LangGraph. Neither is on PyPI or npm yet.

How it fits together

• Event format: what every event carries, and what gets signed.
• Where events are signed: at the first hop outside the agent, so no agent holds a signing key.
• Linking across agents: how events in one agent link to the events in another that led to them.
• Large payloads: what happens to a payload too big for a log entry.

Components

• The proxy on agentgateway and wake-proxy: how MCP and A2A calls are recorded, and how to run the proxy.
• Collector: receives events, signs the SDKs', checks the proxy's and keeps the signed log.
• Signed log: the append-only log, and how anyone can check it with wake-log without trusting Wake.

The brake, the provenance graph, the dashboard and the evidence pack are part of Wake's private platform, and aren't covered here.

Every other page here is copied from the repository as it stood at one commit, named at the foot of the page.