What are AI agent tools? Tools, skills, APIs and MCP explained
A practical guide to the functions, instructions and connection layers that let AI agents take useful action beyond a conversation.
By Scrollport
In brief
An AI agent tool is an executable function the agent can call to retrieve information or change something outside the model. Skills provide reusable instructions, APIs expose software operations, and MCP standardises how an AI application discovers and invokes tools. These layers complement one another rather than describing the same thing.
An AI agent tool is an executable operation the agent can call to retrieve information or change something outside the model. The model decides what to do; the tool performs the bounded action and returns a result the model can use. Tools, skills, APIs and MCP are related, but they solve different parts of that system.
A tool is an executable action
A language model produces text from the information in its current context. A tool gives an agent a defined way to do something else: search a live dataset, verify an email address, generate media or update an authorised workspace. Its definition normally includes a name, a description, a structured input schema and a structured result.
The important word is executable. A paragraph explaining how to verify an email is useful guidance, but it is not a verification tool. A tool has an operation the runtime can invoke and an observable response. OpenAI calls this pattern function calling in its official guide; Anthropic similarly recommends clear, intentionally designed tool interfaces in its guide to writing effective tools for agents.
Tools, skills, APIs and MCP solve different problems
| Layer | What it does | Example |
|---|---|---|
| Tool | A bounded operation the agent can choose and call. | Verify one email address. |
| Skill | Reusable instructions for completing a task well. | Define the evidence gates for building a lead list. |
| API | The software interface that receives requests and returns results. | An HTTPS endpoint behind the verification tool. |
| MCP | A standard connection between an AI application and servers that expose tools. | A remote server exposing apps, discover, inspect, run and wallet. |
A single workflow can use all four layers. A skill tells the agent how to approach the job; the agent chooses a tool; the tool may call an API; and MCP may be the transport through which the AI application discovers and invokes it. The official MCP architecturedescribes MCP as a client-server protocol, not as a replacement for the underlying APIs or the agent’s reasoning.
Capabilities describe outcomes; catalog tools perform them
Scrollport keeps the job separate from the implementation. A category is a broad navigation group. A capability is a provider-neutral outcome, such as email verification. A catalog tool is one specific, independently discoverable, inspectable, priceable and runnable operation supplied by a provider.
That distinction lets an agent search for the outcome first and compare available tools second. Two providers can serve the same capability without pretending that their inputs, evidence or pricing are identical. Explore the live email-verification capability to see the hierarchy in practice.
How an agent finds and uses a tool
Scrollport gives an agent five stable control tools rather than loading the whole catalog into every conversation. apps shows which connected apps are available to the agent, discover searches by outcome, inspect loads one current contract and price, run executes that exact tool, and wallet exposes balance and human-set spending boundaries.
- Describe the outcome, constraints and evidence the job requires.
- Discover a short list of suitable capabilities and catalog tools.
- Inspect the chosen tool’s current schema, access state, example and price.
- Run the bounded operation and use its structured result in the wider task.
This is progressive disclosure: the agent loads detail when that detail can change a decision. Read why Scrollport uses a small control surfacefor the architecture behind the pattern.
Choose the layer that matches the job
Use a model alone when the answer can be reasoned from information already in context. Use a tool when the task needs live data, deterministic execution, specialist computation or an external action. Add a skill when the quality of the result depends on a reusable process, evidence standard or output format. Use an API or MCP according to how the surrounding application stores credentials and transports tool calls.
The layers should make the agent more capable without making its authority vague. A tool remains bounded, its price and access state remain inspectable, and the human still controls connections and spending. To see the working surface, browse the tool catalogor follow the quickstart.