Back to Work

OpenClaw Multi-Agent Ops

OpenAI
Typescript
Node.js
React

This is the personal AI ops system I run day to day. It is built on OpenClaw, an open-source agent gateway, deployed in Docker on an Ubuntu VPS. I did not build OpenClaw itself. I designed and operate everything on top of it: the agent team, their instructions and memory, the scheduling, the internal tools, and the ops scripts that keep it running.

The shape is one main agent and a bench of specialists. Clawvin is the orchestrator: I talk to it over Telegram (text or voice notes) and Discord, and it either answers or delegates. Specialists cover engineering (Patch), QA (Cypress), business development (Scout), research (Alpha), outreach (Iris), content (Nova), plus finance, health, and training agents for personal ops. Each one gets its own workspace with versioned AGENTS, SOUL, TOOLS, HEARTBEAT, and MEMORY files, so instructions, persona, tool rules, and durable memory are files I can diff and review, not hidden state.

Delegation runs through a task board instead of agent-to-agent chatter. Clawvin files a task in Mission Control, a specialist claims it on its next heartbeat, posts progress as comments, and moves it through the columns. Cypress verifies claims against artifacts before work counts as done. Roughly two dozen cron jobs handle the recurring work: briefings, data pulls, reports, backups, and health checks. Mission Control and File Explorer, both covered in their own entries, are the operator UIs I built on top.

The interesting engineering was keeping a system of autonomous agents boring to operate. The gateway moves between pinned upstream release tags through an updater that refuses a dirty tree, rebuilds, health-checks, and rolls back automatically on failure. Restarts go through a script that waits for in-flight turns to finish, validates config, backs it up, and rolls back if the gateway does not come up healthy. A watchdog reaps leaked model processes that used to push the container into OOM, and cron schedules are staggered against a concurrency cap.

Model routing taught me to treat fallbacks as a cost and correctness problem. Agents run on OpenAI models with OpenAI fallbacks. A paid secondary provider is kept out of the chain entirely and only gets added by a gate that confirms a real provider outage. Auth lapses and config errors raise an alert instead, so a broken login cannot quietly route traffic to a metered backup.

Security got real attention. Secrets live only in gitignored env files, config uses placeholders, and gitleaks runs on every push to the config repo. When a self-propagating npm supply-chain worm hit packages that sat in the agent stack's dependency tree, I audited every install on the host (none of the compromised versions were present), pinned the affected dependency, blocked the compromised scopes so any install fails loudly, and added a scanner that runs after every upgrade. Agent-driven installs moved to frozen lockfiles. Commands run against an exec allowlist, and an audit script scans every agent prompt for instructions that would fall outside it and page me for approval.

The biggest lesson: prompts are production code. After one migration, stale URLs and retired workflows lingered in agent instructions and cron prompts and caused drift that took a dedicated recovery pass to untangle. Now there is one documented source of truth per system, corrections have to be written down in a file (and the agent has to say which file), and the whole home directory is a git repo with a written disaster-recovery path.

Capabilities

  • Main orchestrator plus around ten specialist agents
  • Per-agent workspaces with versioned instructions and memory
  • Telegram and Discord channels, including voice notes
  • Task-board delegation through an authenticated agent API
  • Scheduled jobs for briefings, data pulls, reports, and backups
  • Release-tag updates with health check and auto-rollback
  • Idle-aware safe restarts with config validation and backup
  • Outage-gated model fallbacks with alerting
  • Secret scanning, supply-chain scanning, and exec allowlisting
  • Git-backed config with a written disaster-recovery runbook

More Screenshots(2 images)