MCP
MCP is often described as a "USB-C for AI applications": a common way to connect assistants, agents, or IDEs to external systems.
Instead of building custom integrations for each combination of application and tool, MCP defines a common interface between clients and servers.
What it is for
MCP can allow an AI application to access:
- local files;
- repositories;
- databases;
- documentation;
- internal tools;
- ticketing systems;
- search engines;
- APIs;
- specialized prompts;
- workflows.
This improves context management and can facilitate grounding, automation, and agentic work.
Basic concepts
| Element | Function |
|---|---|
| Host | Application where the user works, such as an editor, chat, or AI tool. |
| MCP client | Component that connects the host to MCP servers. |
| MCP server | Exposes tools, data, or resources to the AI application. |
| Tools | Actions the agent can invoke. |
| Resources | Information it can query. |
Relationship with AI agents
An AI agent needs context and tools to act. MCP can serve as infrastructure to provide them in a standardized way.
But MCP does not automatically turn a system into an agent. It only facilitates a connection. Agency depends on the model, the instructions, the ability to plan, the permissions, the sensors, and the oversight.
Risks
Connecting AI to external tools increases value, but also risk:
- exposure of sensitive data;
- excessive permissions;
- execution of unwanted actions;
- prompt injection from external sources;
- untrusted servers;
- loss of traceability;
- dependence on immature integrations.
The practical principle is least privilege: the agent should only access what it needs for the task.
MCP vs. A2A
| Protocol | Main focus |
|---|---|
| MCP | Connecting AI applications to external data, tools, and resources. |
| A2A | Enabling communication and coordination between agents. |
MCP connects agent to tool. A2A connects agent to agent.
Common mistake
Believing that MCP solves context management on its own. MCP facilitates connections, but it does not decide which context is relevant, which permissions are safe, or how to verify outputs. Without governance, it can amplify risks instead of reducing them.
Resources
🏦 Harness EngineeringSkill Arena · Scrum Manager
References
- Anthropic. (2024). "Introducing the Model Context Protocol," Anthropic.
- Model Context Protocol. (2025). "Specification," modelcontextprotocol.io.
See also
Want to advance in agility? You can look for calls for courses and exams, or go at your own pace by becoming a member of the Agile Club. This membership includes exclusive resources, e-learning classrooms, and access to Skill Arena: a space to practice and measure your agile skills at your own pace.