OpenAI Plugins: package examples, portable formats and connection boundaries
Choose a skill, plugin package or MCP service by the capability you need
Compare instruction reuse, package distribution and external-service tools without treating them as competing implementations of the same layer.
What you will learn
- A skill and a service solve different problems
- Choose distribution based on audience
- Evaluate the smallest complete workflow
Before you start
- Basic JSON and directory-path knowledge
- Understanding of skills and external-service permissions
Distinguish source examples from current format guidance and inspect a package without confusing metadata with runtime proof.
Key takeaways
- Skills, packages and MCP services occupy different layers.
- Audience determines the distribution path.
- Choose the smallest system that fully satisfies the task.
A skill and a service solve different problems
A skill captures a repeatable workflow for an assistant. A plugin packages capabilities for discovery and distribution. An MCP service exposes tools or resources that may live outside the conversation. A workflow can use all three, but a small personal writing routine may need only instructions, not a remote server.
Conversely, instructions alone cannot create authenticated access to an external system. A design workflow requiring live Figma data needs the appropriate connection and permission. Installing more instruction files does not repair missing authentication, and adding a service does not automatically provide a well-designed workflow.
Choose distribution based on audience
Local marketplaces support authoring and testing. Workspace publication addresses selected organizational users, while the universal public directory has a separate submission path. Choose the audience before investing in packaging metadata, hosted endpoints and review materials; these channels have different operating requirements.
Use the repository as a set of examples rather than a promise that every package fits your account or host. Its separate API marketplace is a clue that discovery surfaces can differ. Verify the intended installation path and connection support instead of assuming a plugin visible elsewhere must be usable in your environment.
Evaluate the smallest complete workflow
Write one acceptance task and identify the minimum needed instruction, tool and distribution layers. Reject a design that adds a service without a data requirement or bundles unrelated skills only to appear comprehensive. The smaller system is useful when it still produces the full required outcome and supports the necessary review.
Compare candidates using actual output, evidence, permissions and maintenance burden. Do not infer quality from a repository’s weekly ranking or an example’s brand name. This series provides a selection method, not a measured ranking of plugin vendors or a claim that one architecture always wins.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Define the required output and external data needs.
- 2
Assign instruction, tool and distribution responsibilities.
- 3
Confirm host and account support for the chosen path.
- 4
Compare accepted outcomes and maintenance effort.
Copy-ready example
{
"selectionExample": true,
"task": "synthetic meeting-note extraction",
"externalDataRequired": false,
"firstStep": "skills-only trial",
"publicDistributionRequired": false,
"liveTrialCompleted": false
}Frequently asked questions
Do I need an MCP server for every skill?
No. A workflow without external capabilities may remain skills-only.
Does a public example guarantee access in my account?
No. Host, authentication and connection support still need verification.
Sources
- OpenAI Plugins / README.mdSource checked 2026-09-14
- OpenAI Plugins / plugins/figma/README.mdSource checked 2026-09-14
- OpenAI — Package your plugin (checked 2026-09-14)Source checked 2026-09-14
- OpenAI — Plugins usage (checked 2026-09-14)Source checked 2026-09-14