Managed skills
Versioned, server-side instructions your clients' AI assistants fetch live — update the procedure once, and every client's assistant follows the new version on its next use.
What a managed skill is
A managed skill is a procedure you write once and deliver to every client's AI assistant — Claude, ChatGPT, or any MCP-capable agent. The skill's instructions live on your TaskJuice account, not on the client's machine. What the client installs is a thin pointer: a small artifact that tells their assistant where to fetch the current instructions and which of your workflow tools to call.
That inversion is the point. A skill file pasted into a client's assistant goes stale the day you improve the procedure. A managed skill never does: you publish a new version, and every installed copy follows it on the next use. Nothing to redistribute, nothing to version-chase across clients.
Skills ride the same per-workspace MCP connector that exposes your workflows as tools — see AI agents (MCP) for how the connector itself works. A workflow tool is something the assistant can do; a skill is the procedure for when and how to do it.
The lifecycle at a glance
| Stage | Where | What happens |
|---|---|---|
| Author | Settings → Skills | Create the skill's identity (name, permanent slug, description), then write versioned markdown instructions. |
| Publish | Skill detail page | Publishing makes a version active and immutable. Rollback re-activates an old version; it never edits one. |
| Grant | Workspace Admin → Skills | Grant the skill to a client workspace — manually for retainer arrangements, or bound to a product so the grant follows the client's subscription. |
| Install | Client portal → Skills | The client downloads the Claude skill zip or copies the ChatGPT instruction block, once. Claude Code needs no install at all. |
| Evolve | Skill detail page | Publish new versions. Every install follows the newest active version on its next fetch. |
Two readers, two surfaces
Skills have an agency side and a client side, and the docs split the same way:
- You author and grant. Create and publish a skill covers the catalog, versioning, and the advisories that protect installed clients. Grant skills to clients covers delivery, suspension, and usage.
- Your client installs and uses. Install a skill is written for the client persona — the portal's Skills page, and the one-time setup per assistant.
What the client's assistant actually receives
When a client's assistant uses a skill, it fetches the current instructions over the workspace's MCP connector, live. The install artifact itself contains only:
- the skill's name and description (so the assistant knows when to reach for it),
- the workspace's connector address,
- the fetch-then-follow procedure.
It never contains the instructions themselves, a version number, or any credentials. A leaked artifact tells nobody what your procedure is — fetching the real content requires the client's own authenticated connector session, and suspended or revoked grants simply stop serving it.
Plan availability
Managed skills are available on plans with the managed_skills feature. The number of non-archived skills per account is plan-capped (Growth 25, Scale 100) — archiving a skill frees its slot. Skill fetches share the account's daily MCP tool-call budget with every other tool call; see Limits.