agenticloops

Install and run agentic loops — recurring AI agents defined in a single LOOP.md file: a trigger, a set of skills, and a prompt. One file defines it; any harness can run it.

The directory lives at agenticloops.dev. This is the install CLI — the loop-level analogue of npx skills add.

# install a loop onto the harness in the current project (auto-detected)
npx agenticloops install <owner/repo>

# target a specific harness
npx agenticloops install <owner/repo> --harness=5dive

# search the directory
npx agenticloops find research

# what's installed here + global install counts
npx agenticloops list

Commands

Command	What it does
install <owner/loop>	Fetch + validate the LOOP.md, install its skills, pre-flight its requires, and register the recurring job on your harness.
find <query>	Search the public directory.
list	Loops installed on this machine + their global install counts.
update [<slug>]	Re-fetch + re-install (all, or one).

You can also point install / validate at a local path (./my-loop or a LOOP.md) while you're authoring.

How install works

Per the spec:

1. Fetch + validate the LOOP.md against spec v0.1 (required: name, description, a schedule/event trigger).
2. Install skills (skills:). owner/repo/skill paths are fetched directly; bare names resolve against the skills registry; anything the harness already provides is used as-is. An unresolvable skill is skipped with a warning — a loop degrades rather than refusing to install.
3. Pre-flight requires (cli / secrets / mcp / network). This is declare-and-check: missing binaries, secrets, and servers are reported and you're prompted or the install refuses. Nothing is ever auto-installed — binaries, secrets, and MCP servers run code or carry trust, so that consent is yours.
4. Register the trigger on the harness (a recurring 5dive task, a GitHub Actions workflow, a cron scaffold, …).

Harnesses

5dive · github-actions · claude-code · cursor · cron

Auto-detected from the current directory/env. A loop needs scheduling, so a run-only harness (an IDE) can run the agent but can't honor a recurring trigger — the installer warns you and suggests a scheduling harness.

Telemetry — anonymous, opt-out, and exactly this

On a successful install, the CLI sends one fire-and-forget ping to agenticloops.dev/api/install so the directory can show real install counts (like npm download counts):

{ "slug": "ci-analyst", "ts": "2026-06-30T12:00:00.000Z" }

That is the entire payload. The loop slug is a public id. We do not send or store your IP, username, machine id, repo, prompt, or any other field. Updates don't re-count. Failure is silent and never blocks an install.

Opt out any of these ways:

npx agenticloops install <owner/repo> --no-telemetry
AGENTICLOOPS_NO_TELEMETRY=1 npx agenticloops install <owner/repo>
DO_NOT_TRACK=1 npx agenticloops install <owner/repo>   # cross-tool standard

License

MIT. Spec + directory: github.com/5dive-ai/loops.