Patterns
MCP use cases
MCP servers ship for any system you want an LLM to read from or act on. The practical use cases sort into a small number of patterns based on which primitive does the work: Resources for context, Tools for actions, Prompts for templated workflows, Sampling for agentic recursion.
The use-case taxonomy
Document and data context
Primitive: Resources. The model reads streaming content into its context window. File systems, Google Drive, databases, knowledge bases. Anything URI-addressable with content-type metadata.
Action invocation
Primitive: Tools. The model invokes verbs with structured arguments and gets structured output back. Send a Slack message. Open a GitHub PR. Query Postgres with a generated SQL statement. Each verb has a JSON Schema input definition.
Templated workflows
Primitive: Prompts. The server offers parameterized message templates the user can fill in. Code-review templates, summarization templates, customer-support response templates.
Agentic recursion
Capability: Sampling. The server requests the agent to run an LLM completion on its behalf. The server processes a long document by asking the agent's LLM to summarize before returning. Server uses the agent's credentials throughout; it never holds its own model credentials.
Real-world deployments
Across the deployments named in production MCP coverage:
- Software development · coding agents with codebase Resources and Git tool integrations. GitHub MCP server, n8n MCP server for workflow build, GitLab MCP server.
- Productivity · Notion, Linear, Asana, Slack as data sources and action targets. Read tickets via Resources; create or close them via Tools.
- Data analysis · Postgres, BigQuery, Snowflake as Resources plus query-execution Tools. Schema introspection through URIs; query results through structured Tool output.
- Customer service · RAG over policy documents through Resources plus ticket-system actions through Tools. The combined use of historical knowledge and live system access.
The pattern that keeps recurring
Production servers combine multiple primitives in one server to mirror the underlying system's natural shape. A database MCP server exposes schema (Resource), runs queries (Tool), and offers query-builder templates (Prompt). A GitHub MCP server exposes repository contents (Resource), creates PRs and comments (Tool), and offers code-review prompt templates (Prompt). Single-primitive servers exist but multi-primitive servers are the common shape.