Luciq shipped agent skills that arrive with the MCP server rather than being installed, and it is the first genuinely different answer to skill distribution I have seen. The mechanism is a tool: get_luciq_skills, called with no arguments, returns an index of every skill; called with a name, the full guide; called with name and file, a reference file. The skills themselves are procedures over the rest of the surface — “step-by-step guides that orchestrate the other MCP tools for multi-step jobs such as investigating a production signal, producing an app-health readout, configuring or tuning alerts, verifying an SDK upgrade, deduplicating bugs, or auditing PII masking.” The behavioural claim is the one to notice: “your agent does this by itself. Before it starts a multi-step Luciq job it checks whether a skill covers it, then follows that skill.”
What that buys is versioning, and they say so plainly: “now the server carries them. Connect the MCP and your agent has the current set,” with the index refreshing at least daily. Every other skill distribution model in circulation is a file you copy — from a repository, into a directory, at a point in time, and stale from then on. Serving skills as a tool call collapses the gap between publishing a procedure and an agent following it to about a day. The cost is the obvious one: the procedure now requires a live connection and a working credential, so a skill you cannot fetch is a skill you do not have. For observability work, where the procedure changes as the product does, that trade looks right.
The surface underneath is substantial. get_luciq_skills is one of 35 tools across 16 domains — crashes, occurrences, app hangs, APM groups and funnels, bugs, reviews, surveys, session replays, version adoption, AI issues, opportunities, alerts, incidents and documentation search — reached with one-click OAuth 2.0. The catalog carries 13 API pages for Luciq, and the MCP tool names map onto them: the App Health Metrics Export API is the data the app-health readout summarises, Luciq Webhooks is the production signal that starts the investigation, and the iOS and Android SDKs are what the SDK-upgrade verification skill checks.
Luciq scores 52.0, developing on the Kin Score, with contract quality at 61.3 and delegated_identity lit on Agent Readiness — rare, and exactly what one-click OAuth on an agent surface should produce. Agent Readiness is 30.0, agent-ready. Both dimensions this announcement is about read unlit, and the reason is discoverable rather than mysterious: the MCP server’s own overview page names 35 tools and an OAuth flow but never prints the endpoint URL, and it is not resolvable at the obvious hostnames. A server nobody can address is a server no catalog can verify and no agent can reach unaided. The cleverest skill-distribution design of the year is currently gated behind a URL you have to already know.