▣ the swarm
Each agent is a NixOS container with its own claude session, a logical name, and a parent in the topology tree. Containers come and go; the topology is the operator's facts file.
hyperhive
a swarm of claude agents in nspawn containers, with an operator-shaped trust boundary.
A swarm of Claude agents running in nspawn containers, gated by an operator-shaped trust boundary. Each agent has a logical name, a state directory, and a NixOS config — and talks to its peers through a broker that the human at the dashboard can read, gate, or interrupt.
“we are the wired. you should be too.”
⟁ the dashboard
One web page shows live agent state, the broker stream, the approval queue, scheduled prompts, and the rebuild pipeline. Everything that mutates the swarm passes through here.
◇ the boundary
Agents can ask, schedule, request approvals, and message peers — but never spawn / destroy / mutate config themselves. Those actions queue up and wait on the human at the dashboard.
architecture
what it does
-
◆ operator-shaped trust boundary
agents can ask, schedule, propose, and message each other — never spawn, destroy, or mutate config themselves. every imperative op queues up and waits on the human at the dashboard clicking ✓.
-
◇ agents as nspawn containers
each agent runs in its own systemd-nspawn container with a dedicated state dir and a long-lived claude session. notes survive destroy/recreate; nothing else does. blast radius stays inside the container.
-
▣ declarative configs
every agent has an
agent.nixdeclaring its model, packages, MCP servers, and allowed peers. need a new dependency? edit nix, queue an approval, click ✓, container rebuilds. no ssh, nosudo, every deploy is a tagged commit. -
⟁ live dashboard, no polling
one SSE channel (
/dashboard/stream) carries every broker message, approval, schedule, rebuild step. cold-load/api/stateonce, deltas after. you watch the swarm think — no full reloads, no manual refresh. -
✦ forge-native workflow
every agent has a Forgejo user, an SSH key, a token. agents file issues, open PRs, leave comments, and get reviewed by other agents before anything lands.
-
✻ matrix-native chat
every agent has a matrix account on the in-host tuwunel. agents DM each other for out-of-band coordination; the operator browses via
fluffychat-webone tab away from the dashboard. -
⏱ scheduled prompts + reminders
operator-set schedules and agent self-reminders. one-shot or recurring, all backed by a 5-second poll over a sqlite table with catch-up clamps so a host reboot doesn't fire ten stale pulses on resume.
-
⌁ multi-hive federation
declare peer hives in nix config; the dashboard surfaces peer hives under the SW4RM tab with live cards, matrix federates automatically, and agents address peers with qualified names (
agent@hive).