MCP Reference and Use Cases
MCP is the connective tissue of agentic AI: architecture and security matter, but the payoff shows up in deployments, where servers and connectors turn a standard into working automation. This reference collects the common patterns. For the underlying roles and transports, start with MCP Foundations and Architecture.
Three operational models
| Model | Where it runs | Main advantage | Example |
|---|---|---|---|
| Local | IDEs, desktops, private networks | Full privacy, offline access | A local LLM with note and search MCPs |
| Remote (HTTP) | Public, multi-tenant servers | Connects hosted agents to cloud data | A hosted assistant querying a cloud database |
| Hybrid | Local trusted data + remote public APIs | On-prem security with cloud reach | Local CRM MCP plus an online ads-API MCP |
These coexist. A single workflow can orchestrate private and public resources at once — the model doesn’t care where a tool lives, only what it exposes.
Integration matrix
| Domain | Representative server | Key tools | Outcome |
|---|---|---|---|
| Database & storage | Supabase MCP | execute_sql, list_tables, get_advisors |
Query and maintain databases from an agent |
| Web automation | Chrome DevTools MCP | navigate_page, screenshot, performance_trace |
Automated diagnostics, QA, and frontend testing |
| CMS | WordPress MCP | create_post, upload_media, update_theme |
AI-assisted publishing and content automation |
| Creative / 3D | Blender MCP | create_scene, apply_texture, render |
Procedural graphics and asset generation |
| Memory | Context7 Memory MCP | store_context, retrieve_vector, query_memory |
Persistent context between sessions |
| Analytics | Ads-API MCP | fetch_metrics, analyze_campaign |
AI-driven marketing insight |
| DevOps / CI-CD | Pipeline MCP (e.g. Airflow) | trigger_pipeline, monitor_task |
Agents inside CI/CD pipelines |
| Local productivity | Obsidian / SearXNG MCP | read_note, search_web |
Offline personal-assistant automation |
Browser-native MCP
A newer frontier extends MCP directly into the browser. WebMCP lets a website expose tools to a visiting AI agent through navigator.modelContext, so pages become natively “agent-readable” — structured tool definitions in place of fragile screen scraping, with no separate server to install.
Deep dive: Google WebMCP: Direct Agent–Website Interaction.
Worked workflows
Software engineering. Stack: an AI-enabled IDE → model → Chrome DevTools MCP + a source-control MCP. The agent is asked to optimize a page, runs a performance trace through DevTools MCP, identifies bottlenecks, and opens a pull request through the source-control MCP — closed-loop debugging without hand-written glue.
Marketing and analytics. Stack: a hosted assistant → analytics MCP + database MCP. The agent fetches campaign metrics, correlates them against sales data in SQL, and writes a dashboard report through a write-scoped tool — daily insight without manual exports.
Knowledge and productivity. Stack: a local LLM → note + search MCPs. “Add today’s favorite tracks to my music notes”: one MCP fetches the tracks, another enriches metadata, a third writes the note — entirely local and private. For document Q&A over PDFs with tables and diagrams, see MCP RAG Implementation Example.
Data science. Stack: a desktop assistant → a Python MCP server → document store + SQL engine. A stakeholder asks “Q3 spend versus forecast”; the server finds the forecast doc, runs the query, and returns aggregated metrics for the model to narrate — collapsing a multi-hour manual pull into a single request.
Web automation and testing. Stack: a model → a browser MCP + database MCP. Simulate a login flow, capture DOM and network logs, and store benchmarks for regression comparison — continuous validation of production UIs.
SDKs for building servers
| SDK | Language | Fit |
|---|---|---|
| FastMCP | Python | Lightweight stdio/HTTP servers, fast to stand up |
| MCP TypeScript SDK | TypeScript | Node-ecosystem servers (DevTools, database connectors) |
| MCP PHP SDK | PHP | WordPress, Laravel, and custom web platforms |
| Semantic Kernel MCP connector | C# / Python | MCP inside Microsoft agent runtimes |
Common across SDKs: tool schemas, auth helpers, transport abstractions, logging hooks, and test scaffolding.
A composite workflow
An end-to-end “product launch” run shows several servers cooperating:
| Step | Server | Action |
|---|---|---|
| Research | Search MCP | Collect competitor data and trends |
| Storage | Database MCP | Persist structured research |
| Content | CMS MCP | Draft and publish the launch page |
| Analytics | Ads MCP | Pull early campaign metrics |
| Debug | DevTools MCP | Optimize page load and accessibility |
One agent drives the whole cycle, each MCP server acting as a trusted interface at every step.
Emerging enterprise patterns
- Federated directories — central registries of approved servers for corporate agents.
- Fine-grained access scoping — OAuth 2.1 with tool-level policy (see MCP Security and Compliance).
- Event-driven chains — servers triggering successive servers over a message bus.
- Standardized logging — OpenTelemetry trace format across all MCP exchanges.
- Cross-framework orchestration — different agent SDKs interoperating over the shared protocol.
Learning resources
| Resource | What it is |
|---|---|
| modelcontextprotocol.io | Official specification and changelog |
| github.com/modelcontextprotocol/servers | Public registry of community and enterprise servers |
| Supabase MCP server | Remote HTTP implementation with auth and advisors |
| Chrome DevTools MCP | API reference and tool catalog |
| Context7 | Example memory MCP for persistent context |
Takeaways
- MCP standardizes connectivity between models and real systems.
- Its reach spans databases, developer tools, web automation, CMS, and local workflows.
- Modern agent frameworks lean on MCP for cross-system interactivity.
- Open-ecosystem servers demonstrate both versatility and a security model.
- Enterprises are formalizing MCP governance for controlled, compliant access.
