What LoopSkill does — and what it doesn't

Browse and install skills from any agent

Search the public catalog and install a skill into Claude Code, Codex, Cursor, OpenClaw, Hermes, or anything else that speaks MCP. Installation is a pull: your agent fetches the skill. Nothing is pushed to your machine.

223 recorded installs in production.

Run a loop and get an objective pass/fail

A loop executes under enforced bounds and returns a real exit code and duration — not a model's opinion that it went well. This is the part no other skill marketplace does.

Live: run_id, passed, exit_code, duration_seconds returned per run.

Keep separate clients cryptographically isolated

Each client is its own organisation. A credential minted for one org is rejected against another, and the org is derived from the signature — never from a field in the request. So a misconfigured request cannot cross the boundary; it fails.

Two independent paths, both probed against production with a same-org control that must pass first. Mesh credentials: 403 cross_tenant_rejected in both directions. Bundle reads and reconcile: 404 in both directions — indistinguishable from a bundle that does not exist, because a distinct forbidden code would confirm the id is real.

Route a client agent's defect into your private repo

When an agent you operate hits a problem with a skill from your bundle, the report lands as an issue in a private repository you control — not in a public tracker, and not in your inbox.

Private defect routing proven: a report from a client agent lands as an issue in the private sink, verified at GitHub, and the sink is confirmed private by an anonymous request returning 404. The patch-and-redeploy half of the loop has shipped but has not yet been run end to end.

Self-host the whole thing

MPL-2.0. Clone it, docker compose up, and you have the registry, the runner and the catalog on your own hardware. Nothing phones home. Pro is hosting convenience, never a feature gate.

Cold clone verified: healthy container, seeded catalog, README demo runs verbatim.

Fleet convergence reporting

You can see which loop assignments have converged, which are failing, and which have never run. What it does NOT yet do is age a never-run assignment into an alarm — a loop assigned and never fired looks healthy at the top level.

Connector catalog

Thousands of MCP servers are staged behind a review gate. The public catalog is deliberately empty until a human promotes an entry, so you will not find them by searching yet. Staging is not publishing.

Pushing a bundle to a machine

LoopSkill never initiates a connection to your agent or your client's agent. Delivery is pull-only: the agent reconciles itself against the bundle. If you read "deploy to a client" anywhere and understood it as a push, that is not what happens.

A hosted dashboard for your clients

There is no client-facing login, seat, or portal. LoopSkill is a tool for the operator; your clients never touch it. That is a deliberate design decision, not a gap we are racing to close.

Usage-based or per-agent billing

Pricing is a flat subscription. We do not meter agents, runs, or seats, and we have no plans to — skills are plain text and metering them would be theatre.