Open-source always-on AI assistant: event inbox, timers, permissions, outbox
After Anthropic published its article on always-on agents, the author added an always-on layer to waku-agent. An 18-minute video breaks down the design; the core loop is about 95 lines of Python, local-first, and readable. Open source, multi-model support.
Anthropic published an article called "Building always-on agents". After reading it, Shen Sean Chen did a major upgrade to his open-source project waku-agent. The project is a local-first personal AI assistant whose core loop is only about 95 lines of Python—you can read it in an afternoon.
It already had loop engineering, memory, and evals. Now it has an additional always-on layer: an event inbox, timers, permissions, and an outbox—all real code.
Why do you need an always-on agent?
ChatGPT or Claude won't know a task is coming unless you type it into the chat box. Muse, Grok Bot, OpenAI Dots, and Claude Duet are all building personal agents, and the reason is that they are always-on. An always-on agent wakes up when something happens, rather than waiting for you to ask.
System design
In an 18-minute video, the author breaks down four core components:
- Event inbox: A simple database where one row is one event. This is the agent's to-do list.
- Timers: Tell the inbox what to trigger at a future time, such as a briefing every Monday.
- Permissions: Checked on every tool call, and it asks you first when it needs you.
- Outbox: Tells you what it did, or what it needs you to do.
Then he ran it live in Slack. @-mention it, and it wakes up. It refuses to read the inbox in a shared channel. A timer that was set triggers one minute later.
The router for team agents is still on the whiteboard and has not been written into the code.

Why open source?
The author believes open source will help many business owners and startup teams, because you may not want to share all your data with Big Tech. The code in this repo is MIT-licensed (except for the hosted part).
Someone asked: when a timer fires, do permissions check the current state or the state at the time it was set? The author said he is working on it, and his intuition is that it needs to check first, then act. Another commenter suggested adding an idempotency key to each inbox row to avoid processing the same event twice. The author agreed. Another person said the outbox is an underrated part: once side effects are queued, deduplication, retries, and human checkpoints are basically free.
Quick start
To run it directly:
pip install waku-agent
waku
waku dashboard
To read the code or contribute:
git clone https://github.com/ShenSeanChen/waku-agent && cd waku-agent
uv venv && uv pip install -e .
cp .env.example .env
uv run waku
Memory is a SQLite file: ~/.waku/state.db. It supports Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, and others.
The full 18-minute video is in the first reply. The GitHub repo has 2.2k stars.
发布时间: 2026-11-23 05:54