Model Context Protocol (MCP)
Topic
Name: Model Context Protocol (MCP)
Why it matters for FDE roles: MCP appears in AI engineering and FDE listings as a way to connect AI applications to customer tools, files, databases, and workflows without inventing a bespoke integration pattern each time.
Plain-English Definition
MCP is an open protocol for connecting LLM applications to external context and capabilities. An AI app can act as an MCP client, while MCP servers expose tools, resources, and prompts from systems like files, databases, APIs, or internal services.
Where It Shows Up
- Job listing signal: MCP, agents, tool integration, AI-native workflows, enterprise data access, developer productivity.
- Portfolio project connection: Ops Knowledge Copilot could eventually expose records, search, or workflow actions through an MCP server.
- Real customer scenario: A customer wants an AI assistant to read approved internal docs, query operational records, and call a small set of safe workflow actions.
Core Concepts
- Host/client/server split: the AI application hosts the model and connects to MCP servers through clients.
- Tools: callable actions exposed by a server.
- Resources: data or content a server makes available as context.
- Prompts: reusable prompt templates or workflows exposed by a server.
- Capability discovery: clients can inspect what a server offers.
- Security boundary: servers should expose only the data and actions that are appropriate for the user and environment.
Failure Modes
- Treating MCP as a safety layer by itself instead of designing permissions, auth, and approvals.
- Exposing broad filesystem, database, or shell access to an AI workflow.
- Forgetting that tool descriptions and returned content can influence model behavior.
- No audit logs for what tools were called and what data was returned.
- Building an MCP server before the underlying workflow and permission model are clear.
Tiny Practice Task
Sketch a tiny MCP server for Ops Knowledge Copilot with three capabilities: search records, fetch record by ID, and draft a recommendation for human review.
Interview Language
One sentence I could say in an interview:
I understand MCP as a standard interface for exposing tools and context to AI applications, but I would still design the auth, permission, audit, and approval model explicitly for the customer environment.