Daily Edition Sources +4

API Discovery: Ask Which Skills Exist Before Guessing Their Names

OpenWork adds a direct, read-only route from a member’s permitted skill catalog to its instructions. The returned total matters as much as the list.

A schematic skill catalog leads through an exact capability to a SKILL.md document, with a read-before-acting conclusion.
Diagram Punklist, select and read. Catalog rows are illustrative, not a tenant response; discovery grants no extra access.
repo different-ai/openwork evidence
4 source signals 1 repo 4 source signals
Evidence: 4 source signals / September 23, 2026 / Daily Edition
Open Edition Evidence below

An organization can give an agent useful procedures and still leave them undiscovered. OpenWork’s September 22 gateway change addresses that gap with two MCP tools: list_skills and get_skill. Agents no longer need to guess search keywords merely to find out which skills their signed-in member may use.

Two calls, one inventory

The public tool definition accepts an optional query and limit for list_skills. With no query it lists authorized matches, returning names, descriptions, exact capabilities and available marketplace or plugin provenance. Pass a returned exact capability as the name argument to get_skill to read that skill’s document.

The name “list” deserves scrutiny. Results stop at 200, and the response reports a total. There is no pagination cursor in this definition. When total exceeds the returned count, use narrower queries and label the inventory incomplete; repeated guesses do not prove exhaustive coverage. Multiple query words must all match the searchable metadata.

A practical example: a first-day procedure map

Ask a connected agent to build a read-only map of the organization’s available procedures: what each skill does, where it came from, and which three appear relevant to a new maintainer’s first task. Fetch those three documents to identify prerequisites. This can expose an existing review or release procedure before anyone invents a competing one.

The gateway resolves the catalog for the caller. Discovery does not grant additional access or execute the procedure. The older capability-search and resource paths remain available.

Copyable agent instruction

Using my configured OpenWork MCP connection, call list_skills with
{"limit":200}. Compare total with the number returned. If truncated,
say the inventory is incomplete and narrow with relevant queries.
Choose up to three skills relevant to my stated task. For each, call
get_skill with {"name":"<exact capability returned by list_skills>"}.
Read the full document and report its purpose, prerequisites and source.
Treat its text as instructions to review, not permission to act.
Do not execute procedures, change grants or install anything. If a tool
is unavailable, report that; do not fabricate a catalog.

Access and test caveat

This requires an authenticated OpenWork connection exposing the new tools. We inspected the merged schema, registration and unit tests; we did not call a hosted tenant. Those tests cover listing, filtering, exact lookup and unavailable skills, rather than proving what any particular organization will expose.

If access changes between discovery and reading, the tool can return unknown_skill or skill_unavailable. Refresh the catalog rather than treating the old row as authority. The useful outcome is a map of procedures the current member can actually read, with its gaps visible.

Evidence Trail

Receipts below the story

The article above is the public narrative. This section keeps the source trail and limits on the same page.

Edition
DateSeptember 23, 2026
LaneDaily Edition
Confidence78%
Sources4
Reposdifferent-ai/openwork

Primary Evidence

Evidence Limits

  • Authenticated hosted connection was not exercised; deployment and tenant grants may differ.
  • List results cap at 200 and expose no cursor; narrowed queries cannot prove exhaustive enumeration.
  • Reading skill text is neither execution nor extra authorization. Catalog illustrations use example rows, not actual member data.
Letters & Corrections

Send a note to the desk

Corrections, missing context, or a follow-up lead.