The Era of Tool Calls - Why Your Agent Needs Arms and Legs

On The Stage this week, Henry from Smithery said something that broke the whole conversation open:

"We have a bit of a Claude Paradox. You have models that have big brains but no arms and legs."

They can write a sonnet, but they can't check your calendar. They can solve math problems, but they can't ping your team on Slack.

That’s where the Model Context Protocol (MCP) comes in. And if you’re a builder, this is the part where you stop treating AI like a chatbot and start treating it like a teammate.

Here’s what we learned from Henry (Founder of Smithery) and the builders in the room.

1. Tool Calls Are the New Clicks

For the last 20 years, the internet ran on clicks. Humans navigating UIs, finding buttons, making decisions.

In future? They run on tool calls.

Agents aren’t just reading the web anymore; they’re acting on it. When you connect Claude to your tools via MCP, you aren't just getting better answers. You're giving the model the ability to do the work—query a database, fetch a file, execute a script.

Henry’s bet is simple: In the future, agents will do the clicking for you.

2. MCP vs. Skills: The "What" and the "How"

There was a lot of noise in the room about the difference between Claude Skills and MCP. Here’s the fix, stripped of the jargon:

• MCP (Model Context Protocol): This is the capability. It’s the tool itself. "Here is a hammer."

• Skills: This is the methodology. It’s the instruction manual. "Here is how you frame a house with that hammer."

You need both. MCP gives the agent access to the tool. Skills teach the agent the specific workflow you want it to follow so it doesn't hallucinate or break things.

3. Don't Drown the Agent (Progressive Disclosure)

We’ve all done it. You dump 50 tools into your agent's context window and wonder why it gets confused.

The pro move is Progressive Disclosure.

Don't show the agent everything. Show it a high-level menu. If it needs to do deep research, then it loads the deep research tools. If it needs to write code, then it loads the file system access.

Treat your agent's attention span like a human's. Overload it, and it fails. Guide it, and it ships.

---

Try this today:

1. Go to Smithery and find an MCP server for a tool you use daily (Notion, GitHub, linear).

2. Connect it to Claude Desktop.

3. Ask it to do a task you usually do by hand.

It won't be perfect. It might break. Ship the broken version.

The era of human-only clicking is ending. Get your agents some arms and legs.

These conversations happen every Friday at SQ Collective.

Usually over laptops. Sometimes over pizza.

You’re welcome to join the next one.