A local inter-agent communication broker plus a Claude Code plugin that lets AI agents running side by side on one machine — each in its own herdr pane — register under a handle, discover their live peers, and exchange inbox messages (direct messages and request/reply). When mail arrives the broker proactively nudges the recipient through herdr, so a turn-based agent that cannot poll is still told to check its inbox.
The whole system runs for one OS user on one machine, needs no configuration to
start (the first dispatch call auto-spawns the broker), and trusts the OS user
account as its security boundary — no accounts, no secrets.
dispatch is the single artifact that is both the CLI agents/operators run and
the broker it auto-spawns. It is a package-manager dependency invoked bare on
PATH (like git or gh) — it is not bundled with the plugin. Install it with
one of:
brew install herdr-dispatch # macOS/Linux (Homebrew)
nix profile install nixpkgs#herdr-dispatch # Nix
cargo install herdr-dispatch # from crates.ioThe plugin's /register command preflights this and prints the install message
if dispatch is missing.
Installing the plugin gives an agent everything it needs to participate — the
dispatch binary itself stays a separate PATH dependency.
/register <handle> [key=value …]— register this agent under a handle with optional meta. Explicit and manual; there is no auto-register./deregister— cleanly depart, freeing the handle and reaping the inbox.dispatchskill — the agent-autonomous loop: discover peers, send, reply (--reply-to), read the inbox, and ack. It also carries the nudge reaction: on a[dispatch] N unread — run: dispatch inboxline, run the read → reply → ack loop until the inbox is empty.
The broker normally auto-spawns on first use and exits when no agents remain live, but an operator can manage it directly:
dispatch broker start # bring the broker up now (idempotent)
dispatch broker stop # take the running broker down
dispatch broker status # is it running, and its protocol version
dispatch broker prune [--older-than <dur>] # reap retained/crashed records (e.g. 7d)status reports the broker's protocol version so you can tell when a running
daemon has fallen behind an upgraded CLI. prune reaps dead records on your
schedule — death alone frees nothing, so a crashed agent's record and inbox
linger until you prune (or it deregisters). --older-than takes a duration like
45s, 30m, 2h, or 7d.