I got inspired from building clawd2claude. After seeing that my agents are better at communicating with each other directly than receiving instructions from me, I thought that: why should I give instructions to agents when I can have another agent do it for me?
I thought we can emulate real world systems — just like in the office. And that humans might be the bottleneck, where we can't get smart enough to keep up with models to be able to be able to efficiently use them and give good instructions. So... I wanted to build a team of agents that can talk to each other and break down difficult tasks and work on them autonomously.
Kaude is a local multi-agent coding orchestrator. You give it a goal, then watch as three AI agents collaborate autonomously to accomplish it:
- Architect — the planner. Has no tools. It can only think and delegate. It decides what to research, what to build, and when to stop.
- Analyst — the researcher. Read-only access. It explores the codebase, reads files, searches for context.
- Engineer — the builder. Full tool access. It writes code, edits files, runs commands.
You watch them work in a browser UI, just like your manager in your office 😉. If they go off track, you type an interrupt and the Architect adjusts the course to match your expectations.
Work in progress. Will open source later on if everything goes well.