Docs · Sync
How skill sync works
Sync keeps every agent in your fleet running the same skills at the same versions — automatically, without touching your data or your clients' data.
What gets synced
Skill source files
The SKILL.md file and any companion assets bundled in the tarball. These are the instructions your agent reads and executes.
Manifest (versions + signatures)
A signed JSON manifest listing every skill you have access to, its current version, and its SHA256 hash. The manifest is what lets the sync engine know when a skill has been updated.
Install state
Per-agent metadata: which skills are installed, which version, and the install timestamp. Stored locally — the server never writes to your machine; it only serves what you request.
What does not get synced
- Your data — files, credentials, environment variables, outputs, and any artifacts a skill generates at run time. None of that leaves your machine.
- Your client data — if you're on Pro+ and have deployed a bundle to a client's agent, the client's data stays entirely on the client's machine.
- Run-time outputs — anything generated when a skill executes (reports, code, API responses) is ephemeral and local. We never see it.
- Agent conversation history — your chat logs, context windows, and agent state are entirely outside LoopSkill' scope.
How sync works mechanically
- 1
Manifest fetch
At the configured poll interval (default: every 15 minutes), the Recipes meta-skill calls
GET /api/skills/manifestwith your API key. The server returns a signed manifest listing current versions and SHA256 hashes for every skill you have access to. - 2
Diff against local state
The meta-skill compares the manifest against the locally stored install state. If a skill's version or hash has changed — or if a newly installed skill is missing — it queues a download.
- 3
Tarball download from CDN
Updated skill tarballs are fetched directly from the CDN — not routed through the API server. Each tarball is SHA256-verified against the manifest before extraction. Tarballs that don't match are discarded and retried.
- 4
Atomic install
The new version is extracted to a staging directory, then atomically renamed into place. The old version is kept as a rollback snapshot for 24 hours. No partial writes reach your skills directory.
Offline behavior
Skills are local files. If your machine loses internet access, every already-installed skill keeps working exactly as before — execution never calls home. The sync engine simply can't check for updates until the connection is restored.
When reconnected, the next poll fetches a fresh manifest and queues any updates that accumulated while offline. There's no "catch-up" penalty — you just receive the latest versions on the next cycle.
See what's installed on your agents
The library dashboard shows install state per key, including version and timestamp.