AlexOS
Five AI agents that run my workspace in parallel.
A local multi-agent operating system: agents defined in Markdown run as real headless Claude processes, debate a problem over rounds, and feed a shared memory only the server is allowed to write.
chief-of-staff research builder content accountability
| | | | |
+-------- Mastermind round (each sees the others) --------+
|
chair synthesises -> shared memory
agents request actions; only the server writes memory + disk
A Node server spawns five Markdown-defined agents (chief-of-staff, research, builder, content, accountability) as real headless Claude processes. They run concurrently against my actual workspace, with live status streamed over WebSocket to a React front end and cost tracked per run. A Mastermind mode has the panel debate a problem over several rounds, each agent seeing the others' replies, before a chair synthesises the conclusion into shared memory.
The design is security-first, and that is the part I would defend hardest. Agents can only request actions; the server alone writes memory and disk. Instructions are not capabilities, so a persona cannot be talked into corrupting state, because it never had the reach to do so. Scope the tools, not just the prompt.
- multi-agent orchestration
- parallel process management
- WebSocket streaming
- capability-based security
- React + Node
multi-week build
What it actually took. The naive version lets every agent write to disk and it works right up until one goes off the rails. Rebuilding it so the server is the only writer felt like extra plumbing at the time. It is the single decision that makes the whole thing safe to leave running, and it came straight from getting burned by the loose version first.