skills-registry

name	skills-registry
description	Discover what capabilities exist in this org's skills registry — and the MCP bindings needed to call them.
license	MIT
metadata	{"owner":"ITSpecialist111","source":"https://github.com/ITSpecialist111/AgenticSkillstoAgents","readOnly":true}

Skills Registry skill

What this skill does

This skill gives the agent read-only discovery over the organisation's skills registry. The registry is the missing-middle layer between GitHub-tracked capability manifests and the live MCP servers that execute each skill. Use it before invoking any business workflow — it answers:

• "Who in this org can do X?" (e.g. extract fields from an invoice)
• "What categories of work are available?"
• "Is this skill safe to use for the task in front of me?"

The skill does not execute business logic. Each result includes an mcp binding (server name, tool name, transport) that the agent uses to call the underlying skill server directly.

When to use it

Use this skill first whenever a user request maps to a likely capability (invoice processing, PO matching, AP intake, etc.) but you don't already know which concrete tool to call. Typical signals:

• The user names a business document, process, or system.
• The user asks "can you …?" about something outside the agent's built-in toolkit.
• A previous tool call returned a result that needs further enrichment ("now reconcile this against the PO").

Tools

Tool	When to call
find_skill_by_capability(tag)	You know roughly what capability you need. Pass a dotted lowercase tag (e.g. invoice.extract).
describe_skill(skill_id)	You have a candidate skill id and need the full manifest (governance, scoring, preconditions, effects) before deciding to call it. The response includes a payloadFiles list of skill:// resource URIs for the narrative SKILL.md and any asset schemas.
list_capabilities()	You want the full inventory before forming a plan, or to confirm a tag exists.
submit_skill_draft(manifest, payload?)	An agent or human authored a new skill and wants it registered. Opens a PR; a human reviewer must merge before the skill becomes discoverable.

Resources

Each skill's narrative and assets are exposed as MCP resources rather than tool output, so reading them costs no tool slots and no system-prompt context:

• skill://<slug>/SKILL.md — human-readable description and composition notes.
• skill://<slug>/assets/<file> — JSON schemas, examples, scripts.

The <slug> is the skill id with / replaced by - (so finance/invoice-extract → finance-invoice-extract). The URIs are returned in describe_skill().payloadFiles.

Typical workflow

1. User: "Pull the invoice fields out of this PDF and reconcile against the PO."
2. Agent → list_capabilities() to see what tags exist.
3. Agent → find_skill_by_capability(tag="invoice.extract")
   • returns finance/invoice-extract + its mcp binding.
4. Agent → describe_skill("finance/invoice-extract") to confirm the skill's data classification is acceptable for the user's document, then optionally reads skill://finance-invoice-extract/SKILL.md for composition guidance.
5. Agent calls the underlying skill server via the returned mcp binding.
6. Agent → find_skill_by_capability(tag="po.match") and repeats for the matching step.
7. Agent composes the final answer.

What this skill is not

• Not an executor. It returns bindings, not results.
• Not a write-free zone. submit_skill_draft opens a PR, but the PR review is the actual Register gate — nothing lands on main without human approval.
• Not authoritative for credentials. Auth lives on each underlying skill server.