Shaduf.Research preview
OpenAI Plugin Marketplace Guide/Developer guide
Watch: plugins in 65 seconds

A silent, captioned guide to packages, connections and release steps. Real guide screenshots captured 7 October 2026.

The Meeting Evidence teaching source passed static checks. Its installation and behavior remain untested. No plugin installation or invocation is shown.

Transcript and sources

0:00 · One plugin, one useful task. A plugin packages reusable instructions, tool connections, or both. It does not add a new AI model.

0:08 · Three layers. The directory distributes the package. The host uses skills and tools. A connected service supplies authorized data. A listing does not grant account access; provider permissions still apply.

0:20 · Choose the smallest route. A normal prompt may be enough for a one-off task. A skill can package a repeatable procedure. Necessary live data needs an authorized tool connection. Start with readable inputs and an output you can check.

0:30 · Check before connecting. Check the supported host, correct service account, required permissions and actual costs. Access rules can differ between surfaces.

0:40 · Inspect a real example. Meeting Evidence 0.1.0 has a manifest and a skill. It is designed to turn supplied notes into decisions, actions and questions. Its source passed static checks; installation and behavior are untested.

0:51 · Keep release steps separate. Static checks, behavior tests, submission and review, then Publish are distinct checkpoints. A correct ZIP does not prove a working or published plugin.

1:00 · Start with the guide. Make one scoped decision. This video uses the pool's original script and diagrams, with real published-guide screenshots. It does not show a plugin installation or invocation.

Primary documentation checked 7 October 2026: Plugin architecture and Package your plugin. The teaching example and its evidence limits are explained in Build a plugin. Script, diagrams and guide screenshots: Shaduf.

Download MP4 · Download captions

Developer guide · research checked 29 September 2026

OpenAI plugins: the developer guide

A plugin is an installable package of reusable task instructions, tool connections, or both. The universal directory is their shared public catalog for supported ChatGPT and Codex surfaces—not a universal execution or access guarantee.

Plugins overview, Plugin architecture. Checked 29 Sep 2026.

A developer’s path Make each boundary inspectable before the next commitment.
  1. Choose architecture
  2. Build
  3. Test
  4. Secure
  5. Submit
  6. Maintain
Earlier practical studies · original dates and evidence retained

Start from source, not a listing promise

Our original Meeting Evidence 0.1.0 is designed to turn supplied notes into evidence-linked decisions, actions and questions. It has two runtime files and no external account or sending capability. Source syntax/structure and ZIP integrity were checked; behavior, installation, full schema conformance, safety approval and public eligibility were not.

Runtime ZIP · two files · Source + authoring aids

Build a plugin: exact source walkthrough → · Design the missing behavioral tests

Choose your next developer task

  1. Choose architecture

    Decide what needs a skill, a live connection or UI; check the actual host.

  2. Build a plugin

    Inspect the manifest/skill source, download it and check host prerequisites before attempting private installation.

  3. Test and debug

    Run scoped offline checks, repair a packaging failure and prepare separate host evidence.

  4. Security and permissions

    Write the data, authorization and irreversible-action boundary.

  5. Submit and review

    Prepare missing listing and reviewer evidence; keep upload, review and Publish separate.

  6. Distribute and maintain

    Choose local, workspace or public distribution and a versioned update plan.

  7. Business constraints

    Check commerce policy and test a useful-workflow hypothesis—not an earnings promise.

  8. Research

    Read dated studies, their evidence methods and unresolved conflicts.

  9. For AI agents

    Retrieve the right route without dropping exclusions or evidence status.

What public publication would add

Local tests do not publish a plugin. The public path is a complete ZIP, automated checks, final submission, review, approval and an explicit developer Publish action. A skills-only public listing cannot currently gain MCP later; plan the component roadmap first. A public listing does not grant service access or guarantee discovery, approval timing or revenue.

Upload and submit, MCP review and distribution. Checked 29 Sep 2026.

The announcement premise, briefly

The dated evidence supports DevDay 2025 Apps SDK context and a 25 March 2026 Codex plugin milestone. It did not establish a DevDay 2026 universal-directory announcement or today’s directory rollout date. 29 September 2026 is the research check date, not a launch date. The dated report keeps the chronology and disclosed cross-source inference.

DevDay 2025 retrospective, API changelog, ChatGPT & Codex changelog, Plugin UI changelog. Checked 29 Sep 2026.

Need an existing pattern rather than a new package? Inspect Superpowers, GitHub, Canva and Apple Messages.

Search Shaduf

Search published pools, pages, reports, and evidence.