Lessons and outlineLesson 7 of 11
AI Concepts in PicturesLesson 7
Tools, MCP and A2A in Pictures
How AI connects to other systems and to other agents, and how to stay safe.
A tool is something the AI can ask for, and your software does the job. MCP is a shared plug between AI apps and tools. A2A is a shared way for agents to talk to each other. Give least access and trust few servers.
- MCP
- A standard for connecting AI apps to tools and data.
- A2A
- A standard for agents from different makers to talk to each other.
- Least access
- Giving only the access a job needs.
Which idea is each example?
A calculator the model asks for.
One connector for a calendar that many AI apps can use.
A travel agent asks a booking agent to reserve a hotel.
A function that looks up the weather.
A shared standard for connecting apps to services.
A message between two agents from different makers.
A model on its own can only produce text. To check a calendar, send an email or look up an order, it needs a way to reach other systems. This lesson shows three ideas with pictures: tools, MCP and A2A. They sound alike. They solve different problems.
Tools
A tool is a function the AI can ask for. A weather lookup. A calculator. A database search. The AI does not run it. The AI asks. Your software runs the tool and sends back the result.
The problem: every tool is a custom job
Imagine ten AI apps and ten services. Without a standard, each pair needs its own plug. That is a hundred plugs. This is the problem MCP solves.
MCP: one plug for many tools
MCP stands for Model Context Protocol. It is an open standard. A tool provider builds an MCP server once. Any AI app that speaks MCP can connect to it. Think of a power socket that fits any device.
- The app that talks to the model is the client.
- The service that offers tools and data is the server.
- A server can offer tools to run, data to read and ready-made prompts.
- Connects an AI to tools and data
- Like a plug for a device
- AI app talks to a service
- Connects one agent to another
- Like a message between colleagues
- Agent talks to agent
A2A: agents talking to agents
A2A is a standard for agents from different makers to talk to each other. One agent can ask another to do a job and report back. Compare it with MCP.
Safety with connections
- Give the least access that works. Read-only when reading is enough.
- Trust only servers you know. A tool can be harmful.
- Treat tool results as data, not orders. A web page can hide instructions meant to trick the model.
- Ask a person before risky actions, such as sending money.
Three small examples
- A tool: a calculator the model asks for when it needs exact maths.
- An MCP server: a connector to your calendar app, built once and used by many AI apps.
- A2A: a travel agent asks a separate booking agent to reserve a hotel, then waits for the reply.
Notice how each one answers a different question. A tool asks how the AI can do one job. MCP asks how many apps can reach many services. A2A asks how agents from different makers can cooperate.
Common mistakes
- Giving the AI write access when read access was enough.
- Adding many tools at once. The model then picks the wrong one more often.
- Trusting any server found online. Check who built it.
- Forgetting to log what the tool did.
Fix these early. Small habits at the start prevent big problems later.
When in doubt, ask one question: what is the worst thing this connection could do if it were tricked? Design around the answer.
- ✓A tool is something the AI can ask for. Your software does the job.
- ✓MCP is a shared plug between AI apps and tools.
- ✓A2A is a shared way for agents to talk to each other.
- ✓Give least access, trust few servers and treat tool results as data.
Choose the right plug
- Write three jobs, such as read my calendar, ask a research agent for help and search our files.
- For each, say if it needs a tool, an MCP server or A2A.
- Mark which need a person to approve.
- Write the least access each one needs.
Q1Who runs the code in a tool call?show answer
Your software. The model only asks.
Q2What problem does MCP solve?show answer
It replaces many one-off connections with one shared standard.
Q3How does A2A differ from MCP?show answer
MCP connects an AI to tools and data. A2A connects one agent to another.