# Toki

- **Event:** [Built with Opus 4.7: a Claude Code hackathon](https://cerebralvalley.ai/e/built-with-4-7-hackathon)
- **When:** Apr 21 at 12:00 PM – Apr 27 at 2:00 AM (EDT)
- **Where:** Online
- **Team:** [Toki](https://cerebralvalley.ai/u/toki)
- **GitHub:** https://github.com/tokimwc/clawbus
- **Demo video:** https://youtu.be/dS2q9b5zXsU
- **Gallery:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/11

ClawBus is a minimal coordination protocol for multi-agent Claude
Code — a trust layer that makes delegation, approval, and audit
first-class protocol events instead of UI callbacks. It lowers the
trust barrier to multi-agent Claude Code by making every task,
approval, and result inspectable after the fact.

The itch came from running multiple Claude Code sessions across a
5-node cluster for months: every coordination layer I built reached
for workflow engines bigger than needed, while hand-offs still
happened by copy-pasting context between tabs and approving diffs
on vibes.

ClawBus replies with five message kinds (task / result / approval-
request / approval-decision / log), an append-only store with causal
parent links, and an adapter interface four methods long. Three
adapters ship — File, SQLite, Discord; the last lets a human ✅
reaction become an approval-decision without the worker knowing
which transport it's on. Same protocol, different transport,
identical audit trail.

Built on the Claude Agent SDK with Opus 4.7, every file mutation
routes through an approval-request — a fully inspectable message
carrying the proposed diff, severity, and rationale. The gate
doesn't pause on vibes.

MIT TypeScript, 27 tests on CI, three captured end-to-end runs
auditable in five minutes.

---

Markdown version of https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/11. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt