MCP Agents Explained — What They Are and How to Use Them
MCP servers and MCP agents both extend AI capabilities, but they work very differently. This guide explains what agents are, how they relate to the Model Context Protocol ecosystem, and how to use them in real development workflows.
What Is an MCP Agent?
In the MCP Hub context, an "agent" refers to a pre-configured AI assistant persona — specifically, a Claude Code agent. Rather than a running server process, it is a set of custom instructions that specialize Claude for a particular development role or framework.
The MCP Hub directory currently includes 30 Claude Code agents covering frameworks like React, Next.js, FastAPI, Go, Rust, and many more. Each agent ships with:
- A detailed system prompt tuned for the specific technology
- Recommended MCP servers to pair with it
- Example workflows and prompts
Agents vs. MCP Servers — Key Differences
| Dimension | MCP Server | MCP Agent |
|---|---|---|
| What it is | A running process that exposes tools | A custom system prompt / persona |
| Installation | npx or npm install, config entry | Copy-paste into AI client settings |
| Runtime | Runs as a background process | No process — just instructions |
| What it gives AI | Access to external tools and data | Specialized knowledge and behavior |
| Works without AI client | No (needs an MCP host) | N/A (is AI configuration) |
| Best combined with | Any agent or base AI | Relevant MCP servers for the domain |
The most powerful setups use both: an agent persona that gives Claude deep knowledge of a framework, paired with MCP servers that give it real access to the codebase, APIs, and services involved.
How Claude Code Agents Work
Claude Code agents are delivered as custom instruction files. When you load one, it sets Claude's system prompt to include deep context about a specific technology area. The React + TypeScript agent, for example, includes:
- Preferred patterns (hooks over class components, TypeScript strict mode conventions)
- Common anti-patterns to avoid
- Component architecture principles
- Testing strategies (React Testing Library, Vitest)
- Performance optimization patterns
This specialization means Claude does not need to ask clarifying questions about your tech stack or make generic suggestions. It immediately thinks in the patterns that your team uses.
Example Agents on MCP Hub
React + TypeScript Agent
Specialized for modern React development with TypeScript. Knows hooks, context, React Query, Zustand, and Vite. Pairs well with the Filesystem and GitHub MCP servers.
Next.js Full-Stack Agent
Covers the full Next.js App Router paradigm — Server Components, Server Actions, route handlers, and edge deployments. Pairs with Vercel MCP for deployment-aware suggestions.
FastAPI Backend Agent
Python API development with FastAPI, Pydantic V2, SQLAlchemy, and async patterns. Pairs with PostgreSQL and Docker MCP servers for a complete backend workflow.
DevOps Specialist Agent
Infrastructure and CI/CD expert covering Docker, Kubernetes, Terraform, GitHub Actions, and cloud platforms. Pairs with Kubernetes, Docker, and AWS MCP servers.
Security Expert Agent
Focused on application security, OWASP Top 10, dependency auditing, and secure code review. Pairs with Snyk and GitHub MCP servers for integrated security workflows.
How to Use an Agent
Option 1: Claude Desktop (Custom Instructions)
In Claude Desktop, go to Settings > Custom Instructions. Paste the agent's system prompt text. This applies the agent's persona to all conversations in that profile.
Option 2: Claude Code (CLAUDE.md)
For Claude Code specifically, create a CLAUDE.md file at the root of your project. Paste the agent instructions here. Claude Code reads this file automatically and applies the persona to all work in that repository.
# My Project — React + TypeScript Agent
[Paste agent instructions here]
## Project-Specific Context
- Uses React 19, TypeScript 5.4, Vite 6
- State management: Zustand
- Styling: Tailwind CSS v4
- Testing: Vitest + React Testing LibraryOption 3: Cursor / Windsurf / Cline
Each of these editors has its own way to set custom AI instructions — usually a settings panel or a .cursorrules / .windsurfrules file. The agent text works in all of them.
Combining Agents and Servers
The best setups pair an agent (specialized knowledge) with the right MCP servers (real-world access). Some powerful combinations:
- React agent + Filesystem + GitHub + Playwright: Write, review, and test React code end-to-end
- FastAPI agent + PostgreSQL + Docker + Sentry: Build, run, and debug Python APIs with live data
- DevOps agent + Kubernetes + AWS + Datadog: Manage infrastructure with AI that understands your stack
- Security agent + Snyk + GitHub + Semgrep: Continuous security review integrated into your workflow
The Broader MCP Agent Concept
Beyond the Claude Code agents on MCP Hub, the MCP specification itself is adding native agent support. Future versions of the protocol will allow AI agents to act as MCP clients — meaning one AI can delegate tool calls to another AI through MCP. This enables hierarchical agent workflows where a coordinator AI spawns specialist agents, each with their own tool access.
This is an active area of development. The ecosystem is moving toward a world where agents, servers, and tools compose freely across the MCP protocol layer.
Browse all Claude Code agents to find one matching your stack, or read the complete MCP server guide to understand which servers to pair them with.