Understanding Claude Skills, Plugins, and MCP Connectors
A guide to distinguishing between procedural instructions, external data access, and distribution packages.
1Defining the Core Layers
Claude utilizes three distinct layers to enhance functionality: Skills, MCP connectors, and Plugins. Skills provide procedural instructions for tasks, while MCP connectors establish live links to external data sources like CRMs. Plugins serve as distribution packages, bundling these components together for seamless, one-click installation across entire organizational teams or departments.
2MCP Connectors as Access
MCP connectors function as the access layer, linking Claude to external services via the Model Context Protocol. By running an MCP server, Claude gains authorized, live read-and-write capabilities within your existing permission structures. This allows the AI to interact directly with files, calendars, or issue trackers without requiring manual data pasting.
3Skills as Procedural Logic
Skills define the 'how' of a task through written instructions, reference files, or scripts. When a user request matches a skill's description, Claude loads these specific guidelines to ensure consistent output. While skills cannot reach live external data independently, they effectively dictate brand voice, report formatting, or specific review checklists.
4Plugins for Team Distribution
Plugins act as the distribution layer, packaging skills, connectors, and sub-agents into a single installable unit. This approach eliminates the need for individual team members to manually configure each component separately. By bundling these assets, organizations ensure that every user maintains identical setups, significantly reducing onboarding time and configuration errors.
5Layer Interoperability Explained
These components do not exist in a hierarchical power ranking; they are complementary tools. A connector functions without a skill by using Claude's general judgment, and a skill operates without a connector by relying on provided context. Plugins become necessary only when multiple components must be shared and deployed as a unified package.
6Selecting the Right Tool
Match the layer to the specific problem: use an MCP connector for live external data access, a skill for repeatable procedures or templates, and a plugin for team-wide distribution. Starting with the smallest layer that solves the immediate problem prevents unnecessary complexity, ensuring that workflows remain lean and manageable for all users.
7Combined Workflow Example
In a weekly pipeline review, a connector pulls deal stages from HubSpot, while a skill enforces the team's specific review format. Packaging these into a plugin allows a sales manager to deploy the entire workflow to the team in one step, ensuring consistent output regardless of who performs the analysis.
8Marketing and Sales Applications
Marketing teams use connectors for campaign analytics and skills for brand voice consistency, bundled into plugins for writers. Sales teams utilize connectors for pipeline context and skills for follow-up sequences. These configurations are particularly effective for onboarding new hires, as they provide a standardized 'rep toolkit' from their very first day.
9Governance and Security Controls
Anthropic provides robust governance, including automated security scans for public directory plugins and human reviews before listing. Enterprise admins can manage permissions at the tool level, export audit logs via the Compliance API, and view version histories to revert changes if a specific plugin update causes operational issues or errors.
10Permission and Access Management
Connectors respect existing permissions within the host service, such as HubSpot. If a user lacks access to specific pipeline data in the CRM, the connector will not expose that information to Claude. This ensures that security constraints are maintained consistently across the AI interface without requiring a separate, redundant permission system.
11Development Requirements Clarified
Users do not need engineering skills to utilize existing plugins or connectors, as these are point-and-click installations. Writing a custom skill involves plain text instructions, requiring no code. Engineering effort is only necessary when building custom MCP connectors from scratch or submitting new plugins to the official public directory.
12Strategic Implementation Advice
Successful AI integration relies on starting with a single, repeatable task rather than an entire stack. Test skills against defined prompts before wide deployment and pilot plugins with small groups. By focusing on solving specific, manual processes first, teams can build sustainable workflows that scale effectively across the entire organization.
Original article: HubSpot Blog — Understanding Claude Skills, Plugins, and MCP Connectors