← GlossaryGlossary

What is Tool calling?

Tool calling is when an AI agent decides to use a tool — like sending an email, looking up a record, or charging a card — instead of just talking about it.

Tool calling is what turns an AI from a conversationalist into something that can actually do work. The agent gets given a set of tools — functions it can invoke, with defined inputs and outputs. When the agent decides a tool would help, it calls the tool, gets the result back, and uses that result to continue the conversation or take the next step.

Mechanically, the model itself doesn't run the tool — it produces a structured request ("call the calendar tool with these parameters") and the agent runtime actually invokes it. The result comes back to the model, which decides what to do next.

The set of tools an agent has access to defines what it can do. A booking agent without a calendar tool can talk about scheduling but can't actually book. With a calendar tool, it books. The capability gap is huge — and it's why tool calling is the line between chatbots and agents.

A simple example

A booking agent for a hairdresser. Customer says "I need a cut next Tuesday afternoon." The agent calls the calendar tool to check availability, sees Tuesday at 3pm and 4:30pm are open, asks the customer which they want, and when they answer, calls the booking tool to actually create the appointment. The customer never sees the tool calls — they just see a chat that ended with a confirmed appointment.

Why it matters.

Without tool calling, AI is a smarter version of search — useful but inert. With tool calling, AI is a teammate who can actually do work. That's the entire premise of the agent revolution.

For non-technical operators, the practical implication is that you should evaluate AI agents by the tools they have access to, not just the model behind them. An agent on a great model with no useful tools is worse than an agent on an okay model with the right tools.

The risk is misuse. An agent with the wrong tool ("send email" without spam guardrails) or the wrong permissions (a calendar tool that can delete everyone's events) is a hazard. Permissioning and approval queues exist exactly to manage this.

How Squidgy handles it

Tool calling on Squidgy.

Squidgy ships with a library of pre-built tools (calendar, email, CRM, knowledge base, payment, custom HTTP) and supports MCP, so any MCP-compatible tool plugs in directly. You can also write a custom tool for an internal system if you have engineering capacity.

Permission scopes are explicit — when you give an agent a tool, you specify exactly what it's allowed to do with it ("create events but not delete them", "send to existing contacts only, not new addresses"). The platform enforces those scopes at runtime.

Frequently asked

Common questions about tool calling.

Tool calling vs function calling — same thing?+

Yes — different names for the same mechanism. Function calling is the older OpenAI term; tool calling is the more general term. They both refer to the agent invoking a defined function with structured inputs.

How does the agent know which tool to use?+

The model is given descriptions of each tool (what it does, what inputs it needs, what it returns). It picks based on the user's intent and the conversation context. Strong tool descriptions are the difference between agents that pick well and agents that pick badly.

Can the agent use the wrong tool?+

Yes, especially with weak tool descriptions or many overlapping tools. Mitigations: keep the tool list focused, write clear descriptions, scope permissions tightly, and run evals that catch wrong-tool errors.

How do I add a custom tool for my business?+

On Squidgy, two ways. If your service exposes an MCP server, plug it in — done. If it doesn't, write a small wrapper (or have us build one) that exposes the function as a tool the agent can call.

Build your own AI agent.

No code. Hands-on onboarding from the team in your first cohort.