Skip to content
LoopSkill
Open-source toolkit for people who run agents for other people

Deploying AI agents for clients is a job now.
This is its toolkit.

From your first installed skill to a paying client's self-maintaining fleet — one open-source platform for the whole path. Pick the station you're standing on.

What are your agents missing? One command scans your local Claude/Codex/Hermes/OpenCode installs and diffs against the catalog.

Open source (MPL-2.0) · Free tier · Self-host the whole thing · Cancel anytime

The path · click a station

You have an agent. Give it a supply chain.

Day 1 costs nothing and needs no account. Point your agent at the registry, and it can search, install and run vetted skills from inside its own conversation — one MCP surface instead of forty copy-pasted markdown files.

A · the one-liner

Tell your agent this. That is the whole install — no package to add, no key to paste.

"Install the loopskill skill from app.loopskill.io/skill" Per-agent install guide — Claude Code, Codex, Cursor, Cline, Hermes, OpenClaw

B · not ready to sign up? Look at what you already have

loopskill import and loopskill diff read the skills already sitting in ~/.claude, ~/.codex, ~/.cursor and ~/.hermes, and show you exactly how two machines have drifted. Zero network calls — enforced structurally, not promised: the network-capable module is one the offline commands never import, and a test breaks socket.socket for their duration to prove it.

Getting the CLI
agent session
$ you
Install the loopskill skill from app.loopskill.io/skill please
$ agent
→ Fetching SKILL.md from app.loopskill.io/skill
→ Wiring LoopSkill MCP — run loops · search · install · bundles
✓ LoopSkill installed. The runnable registry + 99,000+ skills — ready.
Just tell me what you need.
$ you
Run the secret-scan loop on my repo before I push
$ agent
→ Running secret-scan-loop under enforced bounds (bounded confinement)
→ Scanned working tree · exit 0 · no leaked credentials found
✓ Loop passed in 1.2s. Safe to push. Verdict logged to the registry.
Works in Claude Code, Cursor, Cline, OpenClaw, Hermes & Windsurf — any MCP-capable agent.

What day 1 actually buys you

  • · One search surface over this catalog and 99,000+ skills indexed across federated registries — we link to those, never rehost them.
  • · Skills that ship a runnable loop: press run, get an objective pass/fail under enforced bounds. Not a model's opinion.
  • · Nothing to un-pick later. Everything below is the same platform, no migration.

Hand them one line. Keep your name on it.

A bundle is the unit of client work: the skills that client's agent should have, pinned at exact versions. You edit the bundle; their agent pulls. There is no push — the control plane cannot reach into a client machine, which is precisely why a client will let you near theirs.

  1. step 01

    Compose the bundle

    Pick skills, pin versions, name it after the client. Or fork someone else's public stack and start from a working set instead of a blank page — preview the diff before you claim it. Compose one →

  2. step 02

    Scope a key to that bundle, and nothing else

    A bundle-scoped key (cbt_…) reaches exactly one bundle. If a client leaks it, they leak that bundle — not your catalog, not another client. Key scopes →

  3. step 03

    Send the install line

    One line, branded-neutral, key already baked in. Their agent installs and then reconciles on its own schedule — no dashboard for them to babysit, no LoopSkill branding on their side. Delivery guide →

Before you run someone else's code on a client's machine

  • · Skills can run in an ephemeral, syscall-audited sandbox with a network allowlist.
  • · Anonymous contributions land as drafts and need a human reviewer before promotion.
  • · Packs are content-addressed: the same source rebuilds byte-identical, sha256 verifiable.
The trust stack, in detail

Your second client is where it breaks.

One client, one agent — easy. Two clients, five agents, and nobody knows which version of which skill runs where. A bundle lock is an immutable revision; every agent reports a three-way verdict against it — declared vs installed vs extras. Drive the matrix below.

Fleet + locks guide
DEMO · simulated fleet lock rev 7
2 of 5 agents converged
Simulated drift matrix. Rows are agents across two example clients; columns are the skills a bundle lock declares. Every cell reports the three-way verdict — declared, installed, extras — the same verdict the fleet routes return.
agent copywriting declares 1.0.0 humanizer declares 2.1.1 agentic-os declares 1.0.0 Super Memory declares 1.0.2 ruthless-mentor declares 1.0.1 action
acme-prod-01 Acme · claude code
acme-prod-02 Acme · codex cli
northwind-01 Northwind · cursor
northwind-02 Northwind · hermes
solo-laptop You · claude code

Hover, focus or click any cell to inspect it — or hit Converge on a drifted row.

The state above is simulated so you can drive it without an account. The skills named are real catalog entries — click one. Your own matrix lives behind sign-in: Your Library → Fleets · how locks and the three-way verdict work. Known gap, stated rather than hidden: an assignment that has never run once still reads healthy at the top level — see what LoopSkill does and does not do.

Immutable lock revisions

Edit a bundle and a new revision is minted. Every agent's verdict is computed against a specific revision, so "which version did they actually get" has an answer, not a guess.

Agents enrol themselves

An agent proves possession of an Ed25519 key and registers itself. No OAuth screen, no session, no human clicking through a console on a client's laptop. How →

Convergence is a poll, not a push

Nothing is pushed to a client machine and nothing is real-time — an agent reconciles on the schedule you install. That is a security property, and we would rather state the interval than call it "instant".

Run a Fleet — the curriculum for this station

A multi-agent setup with shared memory, loop-safe Discord coordination, and per-client reporting.

Open the track

The runbook you already wrote is the product.

Three ways the work you're already doing turns into revenue — none of them require you to stop doing client work.

01 · publish

Turn a workflow into a skill

Describe the runbook in plain language and one call scaffolds a SKILL.md draft. It goes through a quality gate and a human review before it's listed — the same gate everyone else's does.

Creator workflow

02 · refer

50% recurring, then 30%

Refer a Pro subscriber and earn 50% of every invoice they pay for your first 50 referrals, 30% after — monthly, for as long as they stay. Paid out via Stripe Connect.

Referral terms

03 · hand it off

Or don't run it yourself

If a client wants the outcome and not the operations, WiseChef runs the same stack as a managed service. Same skills, someone else's pager.

Publish or get help

What it costs to run this

The whole platform is MPL-2.0 — clone it, docker compose up, run every capability on your own hardware with no account and nothing phoning home. Pro is hosted convenience, never a feature gate. See the ladder →

99,000+
Skills indexed across federated registries
61
Curated & signed in this catalog
pass / fail
Every loop run, objective + bounded

Works with every
major AI agent

Claude Code Cursor Codex Hermes OpenClaw Cline Zed

Before you bet a client on this

We publish what this does not do.

You are about to put this between you and someone who pays you. A marketing page that only lists wins is not enough information to make that call, so there is a page that lists the gaps — known blind spots, half-shipped surfaces, and the exact probe behind every "verified" claim.

Example: the never-run blind spot

The fleet view shows which assignments converged and which are failing. It does not yet age a never-run assignment into an alarm — a loop assigned and never fired looks healthy at the top level. That is on the scope page, not buried in a changelog.

Every number on this page is live

The counts above are fetched from public endpoints when this page is built, and any stat that cannot be fetched is omitted rather than guessed. Check them yourself: the API reference lists the endpoints they come from.