Human Note: This is a snapshot, this was developed specifically for working on one other project: Cheyss. It does not convey my impressions of the experience. *
CCC Changelog
CCC is a multi-agent orchestration system. It manages teams of Claude Code instances — giving them identities, chat rooms, task queues, and a shared workspace — then keeps them productive with minimal human intervention.
Week 1 (Feb 10–11) — Built the thing
Started from nothing. By end of day 2 there was a working chat system, a delivery daemon that nudges agents about new messages, tmux-based session management, and a terminal dashboard for the operator to monitor everything. Agents could be added, removed, and would auto-start on launch. The whole system runs from a single workspace directory.
Week 2 (Feb 12–16) — Made agents smarter about context
Agents now wake up knowing what work is available (from the issue tracker) and get periodic reminders about their responsibilities. Added a token ring protocol so room-wide messages get orderly responses instead of everyone talking at once. The operator’s lead agent gets pinged periodically to check on the team. Config moved to a proper TOML file so behaviour is tuneable without code changes.
Week 3 (Feb 17–23) — Google Sheets pipeline & verification
Added a full task pipeline through Google Sheets — the operator drops tasks in a spreadsheet, agents pick them up, and completed work goes through a verification step before being marked done. Built an agent creation wizard in the dashboard. Stored agent personalities and responsibilities in the database so they persist across restarts.
Week 4 (Feb 24–26) — Naming cleanup & stability
Fixed a pervasive bug where agent names with hyphens vs underscores caused duplicate rooms, lost messages, and broken lookups. Migrated all data to a canonical form and extracted a shared normalization function. Removed several features that weren’t working reliably (manual suspend/resume, session checkpoints).
Week 5 (Mar 2–5) — Web UI & nudge intelligence
Replaced the terminal-only interfaces with a web UI that includes live terminal tabs for each agent. Made the system smarter about when to interrupt agents — it now checks whether they’re actually idle before nudging, instead of blindly delivering every message.
Week 6 (Mar 12–13) — Remote agents
Agents can now run on other machines. A single ccc bootstrap command SSHes into a remote host, installs prerequisites, sets up the project, and starts Claude — fully turnkey. The dashboard discovers and monitors remote agents alongside local ones. Also stripped out the handoff system (session brain-dumps that were never actually used since claude --continue handles session continuity natively).