What is the Model Context Protocol (MCP)?

Model Context Protocol (MCP) is an open standard that defines a consistent, secure way for AI applications and agents to access external tools and data sources. In short, MCP is like a USB‑C for AI: a universal connector between LLMs and the real-world systems where your data and actions live.

What is an MCP server?

An MCP server is the bridge that exposes tools (actions) and resources (data) — such as REST APIs, databases, or local files — to MCP clients (agentic apps, IDEs, or chat interfaces). The server defines the capabilities and schemas; the client discovers and safely invokes them.

MCP vs RAG vs function calling

Bottom line: RAG retrieves text; MCP provides structured tools & actions with standardized discovery, permissions, and schemas.

Architecture & core concepts

• MCP client: the AI application/agent (e.g., desktop app, IDE, chat UI)
• MCP server: component that exposes tools/resources via the protocol
• Tools: parameterized actions (e.g., create_issue(repo, title))
• Resources: readable context (e.g., files://specs/*.md, db://analytics/sessions)
• Sessions: scoped interactions with lifetimes, policies, and logs

{
  "tool": {
    "name": "create_issue",
    "input_schema": {"type":"object","properties":{"repo":{"type":"string"},"title":{"type":"string"}}},
    "output_schema": {"type":"object","properties":{"url":{"type":"string"}}}
  },
  "resource": {
    "uri": "db://analytics/sessions",
    "schema": {"type":"array","items":{"type":"object"}}
  }
}

Practical use cases for MCP servers

• Engineering productivity: repositories, CI/CD, issue trackers
• Data & analytics: query warehouses, generate charts, schedule jobs
• Customer ops: CRM updates, ticket summarization, outreach drafts
• Knowledge ops: read knowledge bases and update runbooks
• Security/IT: run scoped diagnostics, fetch logs, open incidents

Quickstart: build a basic MCP server

1. Pick an SDK: Start with official TypeScript or Python server SDK.
2. Define tools/resources: Focus on high-value actions with clear schemas.
3. Guard access: Use API keys/OAuth, per-tool scopes, and allowlists.
4. Test with a client: Connect via an MCP-enabled app/agent and verify flows.
5. Harden: Add timeouts, rate limits, sandboxing, and injection scanning.
6. Deploy: Ship behind a gateway; monitor logs and success/error rates.

Security risks & defenses (must‑read)

MCP unlocks write actions — which means threat modeling is essential. Start with least privilege and explicit user consent for sensitive tools.

• Indirect prompt injection: Treat untrusted content as hostile; scan inputs and add content provenance checks.
• Tool poisoning: Allowlist servers and signed tool manifests; pin versions and hashes.
• Credential leakage: Use short‑lived tokens; never echo secrets back into model context.
• Overbroad scopes: Give each tool the minimum set of permissions and rate limits.

FAQs

References

• Anthropic — Introducing the Model Context Protocol
• ModelContextProtocol.io — Introduction & FAQ
• Anthropic Docs — MCP
• GitHub — Model Context Protocol org
• AWS Blog — Unlocking the power of MCP
• Red Hat — MCP security risks & controls
• Microsoft — Protecting against indirect prompt injection in MCP
• The Verge — Anthropic launches MCP
• Thoughtworks — Beneath the hype: MCP