OlympusMulti-agent DevOps

One human, a coordinator, and a team of LLM specialists driving real infrastructure — tool-gated, human-approved, fully audited.

What is Olympus?

Quick start

Live demo ↗

Group-chat coordination

A main coordinator runs the conversation and dispatches subtasks to specialists (sysadmin, programmer, terraform, ansible, hpc), narrating its reasoning as a live thinking trace.

Safe by construction

Every tool is gated to its agent's declaration; destructive ops require human approval; a self-protection policy stops Olympus managing the cluster it runs on. Everything is append-only audited.

Learns & undoes

Memory retrieval + feedback on past runs, per-verb rollback that re-prompts approval, and per-turn (and per-agent) cost telemetry.

Extensible over MCP

Graft any third-party MCP server (stdio or HTTP) onto a named agent — gated and audited like native tools. NetDB (IPAM/DNS/DHCP) is the production example.

Two interfaces

A CLI that routes a task to one specialist, and a React dashboard group chat with auth, approvals, capabilities consoles, an in-browser terminal, and per-user accounting.

Reproducible deploy

Self-host on Proxmox/bare-metal, or clone the sandbox and stand up the whole stack on AWS (kubeadm + Helm + TLS) with one command.