@crewlet/loopany

The Loopany daemon - runs on your machine, connects to a Loopany server, and executes your scheduled agent loops locally via your own coding agent.

Loopany is BYOA (bring your own agent): the server schedules, stores, and notifies, but never runs an LLM or executes your code. This daemon is the execution half - it polls the server for due runs, spawns the loop's coding agent (Claude Code, codex exec for a codex loop, or the grok CLI for a grok loop) in the loop's working directory, and reports the results back.

Requirements

• Node.js >= 22
• Claude Code installed (claude on your PATH), the OpenAI Codex CLI (codex on your PATH) for a codex loop, or the xAI grok CLI (grok on your PATH) for a grok loop
• A Loopany server - its dashboard gives you the server-url and one-time connect-key used below

Install

npm install -g @crewlet/loopany
# or run ad-hoc:
npx @crewlet/loopany --help

Connect your machine

loopany up --server-url <url> --connect-key <dk_…>

up is idempotent: it registers this machine (first time), stores the credentials under ~/.loopany/, and spawns a detached daemon if none is running. It also refreshes the user-scope loopany skill, the SessionStart hook, and the loopany PATH shim. After that, loopany up alone reconnects.

Commands

loopany                 Show the content-first HOME: this machine's live loops +
                        recent runs (the poll loop moved to `up --foreground`).

Setup
  up [--foreground]       Connect this machine / ensure its daemon is running
                          (idempotent; refreshes the loopany skill, the SessionStart
                          hook, and the `loopany` PATH shim). --foreground runs the
                          poll loop attached in this terminal instead of detached.
  new --json '<config>'   Create a loop from an inline JSON config (--json - reads
                          stdin). --dry-run validates + previews, creates nothing.
  setup hooks [--remove]  Install/refresh the SessionStart hook that lands the home
                          view as ambient context each session (--remove uninstalls).
  skill [status|install]  Manage the loopany agent skill install (user scope by
                          default; --project installs into the current directory).
  update                  Update this machine's daemon to the version you invoked
                          (run via npx @crewlet/loopany@latest update): stops the
                          running daemon, starts the new one, refreshes the
                          skill/hook/shim.

Management
  status                  Is the daemon running? Show pid + server connection.
  down                    Stop the detached daemon started with `up`.
  show [<id>]             Show a loop's full editable config + recent state (the
                          device credential inspects any loop on this machine).
  log [<loop>]            Show a loop's recent runs (status, metrics, session id).
                          --transcript/--full inlines transcripts; --json for machines.

Interactive
  loops [--fields a,b]    List your loops (default columns id/name/cron/enabled/
                          nextFire; --fields adds timezone/notify/model/goal/
                          taskFile/runs/lastOutcome; --json for machines).
  edit <id> --json '<obj>'  Edit a loop (JSON-only + --workflow-file/--ui-file/
                          --schema-file; --dry-run previews before/after).

Run loopany --help for the full usage text, or loopany <verb> --help for a single verb's concise usage (prints and exits, running no side effect - safe to inspect foot-guns like update or down).

How it works

The daemon polls the server over HTTPS - no inbound ports, no websockets. While idle it opts into a bounded server-held long-poll so a due run dispatches almost instantly; with a run in flight it keeps a short poll so progress keeps flowing. When a run is due it executes the loop's coding agent in the loop's own folder, live-syncs that folder's files back to the server (secrets and junk like .env*, node_modules, .git, .worktrees, and build/tool caches are never sent), and reports the outcome. The loop folder is a synced content home (reports, state, ui, small artifacts) - not a scratch workspace: heavy work products (a repo clone, a git worktree, build output) belong outside it, and the daemon defensively caps how much it syncs per loop (LOOPANY_SYNC_MAX_FILES / LOOPANY_SYNC_MAX_BYTES) so a stray checkout can never flood the sync. Your code and credentials stay on your machine.

The package also bundles the loopany agent skill, which teaches a coding agent how to author and evolve loops; loopany up (and loopany new) install it at user scope for every coding agent loopany knows about (Claude Code ~/.claude/skills/loopany/ and Codex ~/.agents/skills/loopany/ today) automatically, so any loop on this machine can discover it. Run loopany skill status to see each agent's install location.