What Are Plugins? Skills, MCP, and Distribution
The instructions and capabilities a plugin packages
A plugin is an installable, distributable package that can bundle skills, connections, and other capabilities an AI uses. A skill describes how to perform a particular task; a plugin packages those instructions so other people can install them. It can include connections to external services, but connections are optional.[1][2]
What you’ll learn
- Explain the roles of skills, MCP, and plugins
- Use a proposal-writing example to distinguish reusable instructions from package distribution
- Check plugin installation separately from permission to access data
Skills provide instructions, MCP connects, plugins distribute
A skill packages reusable steps, decision criteria, and output formats. Its role stays the same when it is included in a plugin.
MCP (Model Context Protocol) is a shared set of rules for connecting AI to external data and capabilities. An MCP server follows those rules to provide functions such as file search and information retrieval. A plugin can include these connections.[1]
A recipe and a cooking kit offer a useful comparison. The recipe corresponds to a skill, and the kit that packages instructions and tools corresponds to a plugin. In practice, this separates the unit for designing how work is done from the unit for distributing it. The analogy has a limit: installing a real plugin does not replace signing in or otherwise confirming who is using the connected service (authentication), or getting permission to read or change its information.[3]
Consider a proposal that tells a customer what will be provided and at what price. The choice depends partly on who supplies the customer’s requirements, budget, and other information the AI needs. Compare what each mechanism reuses rather than how many features it has.
| Aspect | Skill | MCP | Plugin |
|---|---|---|---|
| Main role | Provide steps and decision criteria | Connect data and capabilities | Package capabilities for installation and distribution |
| Proposal-writing example | Structure and review criteria | Connection that retrieves customer requirements and budget | Package containing writing instructions and required connections |
| What it does not determine on its own | Access rights in connected services | Workflow and completion criteria | Which data and actions the authenticated account can access |
Choosing skills and plugins for proposal writing
When a person can supply the customer requirements and budget, a skill can standardize the proposal structure, price checks, and review criteria. At this stage, evaluation checks whether the same input produces a draft containing the required items.
If the workflow needs to retrieve information from a service storing customer requirements and budgets, or from file storage, a plugin can package the connections with the writing instructions. This is an example configuration, rather than a claim that a particular plugin supports every service or action. Retrieval depends on the selected connection and its permissions.[3]
A plugin can also help without external connections when multiple skills should be distributed together, or when a workflow needs to be shared in an installable form.[2] The decision depends on whether the need is to organize instructions alone or also to define a package for distribution.
Plugin installation and access permissions are separate
Making a plugin available and granting access to connected information are separate steps. OpenAI distinguishes plugin availability, MCP connection access, permitted actions, and the permissions of the account authenticated with the service.[3]
For example, installing a file-search plugin does not let an account read documents it lacks permission to view. Reading information and changing it are also distinct capabilities. The ability to create a draft does not establish that the workflow can send it or update an external service.
Supported capabilities vary by product and by the app or screen where the plugin is used, even within the same package. OpenAI and Anthropic both provide plugins that package skills and connections, but each defines which capabilities run in which environments.[4][5] Check the candidate package, the app or screen being used, and the required data and actions before choosing it.
References
- OpenAI, Skills & Plugins
- OpenAI, Build skills
- OpenAI, Plugin controls
- OpenAI, Plugins
- Anthropic, Use plugins in Claude