# AI agent tools: the data and APIs an agent needs

The AI agent tools that matter most fetch real data. A plain map of the jobs agents call tools for, and how to connect them without a dozen keys.

## Summary

Lists of AI agent tools usually mix three different things: frameworks that run an agent, models that power it, and the tools an agent calls to do work. The third group decides whether an agent is useful. A model with no tools can only talk. Give it a search tool and a lookup tool and it can answer with facts.

This post covers the tools in that third group. It sorts them by the job they do, shows what to ask of each one, and explains how to connect them so the agent does not carry a dozen keys.

## What are AI agent tools?

An AI agent tool is a function the agent can call to act on the world outside the model. The agent reads the tool's name, description and inputs, decides to call it, and gets a result back. A tool can search the web, read a page, find an email, query a database or write a record.

Most agent tools sit behind an API. The agent does not call the API directly. A tool definition, often delivered over MCP, tells it what the call needs. If the protocol is new to you, [what is an MCP server](https://looot.ai/blog/what-is-an-mcp-server) explains it.

## What jobs do agents call tools for?

Group tools by the job, not by the vendor. Jobs stay stable while vendors change, and the same job often has several providers.

| Job | What the agent gets back | Example capability in the catalog |
|---|---|---|
| Search the web | Ranked results with titles and links | google.serp.organic, web.search |
| Read a page | Clean text or markdown | web.scrape.markdown |
| Find a person or company | Names, titles, domains, emails | people.email.find, company.enrich |
| SEO data | Keyword volume, backlinks, rankings | google.keywords.volume |
| Social data | Posts, profiles, comments | reddit.post.search, x.post.search |
| Local data | Places, reviews, phone numbers | maps.place.get |

Every row is a read. Writing to other systems, such as a CRM or an email tool, is a fourth kind of tool and needs more care, because a mistake there changes real records.

## What should I check before I add a tool?

Five checks save time later.

1. **Inputs.** Does the agent know every field it must send? A tool with vague inputs makes the model guess.
2. **Output.** Does it return the fields you need, in a shape the agent can read? A result that is a wall of text wastes context.
3. **Price.** Is the cost per call known before the call? Without it you cannot cap a run.
4. **Failure.** What does an empty result look like, and is it billed? A tool that returns nothing and charges anyway needs a guard.
5. **Auth.** Where does the key live? Keys inside a prompt or a repo leak.

When a tool fails check 3 or 4, the fix is a rule in the prompt: stop at a limit, and report empty results instead of retrying.

## How many tools should an agent have?

Fewer than you think. Each tool definition costs context, and a model that sees hundreds of tools picks worse than one that sees a few. The usual advice is to give the agent only the tools the job needs.

The catch is that real jobs touch many providers. A lead list needs search, enrichment and verification, each from a different vendor. If every provider is a separate tool, the list grows fast.

The way out is to hide the long list behind a small interface. The agent gets a few tools that search a catalog and run an endpoint, so the number of tools it sees stays at a handful while the reach stays wide. [MCP gateway for agents](https://looot.ai/blog/mcp-gateway-for-agents) describes that design.

## How does looot give an agent tools?

looot is one connection that gives an agent access to data and tool APIs. The agent holds a token. It has a short set of tools, including discover, inspect, run, runs, balance and top_up. It asks discover for a job in plain words, reads the price and inputs with inspect, then runs the endpoint.

Pricing is per call and shown before the call. If one provider returns nothing, the agent can try another for the same job, following your instruction. The [public catalog](https://looot.ai/public-catalog) lists every endpoint with its price, so you can read what the agent can reach before you connect.

The table below shows the live price of three of the jobs from the earlier table.

Price per call for common agent tools: live prices per call are read from the catalog when the page loads, see https://looot.ai/blog/ai-agent-tools-for-data-and-apis.

- Google organic results
- Find a work email
- Page as markdown

## Which tools should I start with?

Start with the job you already do by hand. If you spend an hour a week copying company details into a sheet, start with the enrichment tools. If you check search results for a topic, start with search. Add a second job only after the first runs cleanly.

For a worked example of enrichment, see the [data enrichment use case](https://looot.ai/use-cases/data-enrichment). It shows the steps an agent follows from a list of domains to a table of contacts.

## How do I test a tool before I trust it?

Run it on three inputs you already know the answer to. One easy case, one with an unusual input such as a company with a long name, and one that should return nothing. The first checks that it works. The second shows how it treats awkward data. The third shows what an empty result looks like and whether you were charged for it.

Write down what you see: the fields returned, the time taken and the price. If the empty case is billed, put a rule in the prompt that limits retries. If the unusual case returns wrong data, note it, because the agent will hit that case in a real list.

## What about tools that write to other systems?

Tools that create records, send messages or change settings are different from tools that read. A wrong read costs a call. A wrong write changes something real and may be hard to undo.

Treat write tools with three extra rules. Have the agent show what it is about to write and wait for approval. Start with a test record or a sandbox if the system has one. And log every write so you can find what changed. Most of the value in a first agent comes from read tools, so add write tools once the reads are solid.

## Try it: a prompt for your agent

This prompt asks the agent to list the tools it has and map them to jobs. It costs nothing, because it only discovers and inspects.

If step 3 returns a price for each job, the connection works and you know what a first run would cost.

## What it cannot do

- It does not cover every API. If an endpoint is not in the catalog, the agent cannot call it through looot.
- It does not choose the right tool for a bad task. A vague instruction gives a vague result.
- It does not run jobs on a timer. Repeating a job needs scheduled runs, not available yet.
- It does not guarantee results. The catalog shows a measured success rate only where one is published, so check the endpoint.

Last checked 2026-10-03 against the live catalog and the looot skill file. Author: the looot team.

## Questions

### What are AI agent tools?

AI agent tools are functions an agent can call to act outside the model, such as searching the web, reading a page, finding an email or querying a database. Each tool has a name, inputs and a result the agent can read.

### What are the best tools for AI agents?

It depends on the job. Search and page reading cover research. Enrichment tools cover lead data. SEO endpoints cover keyword and backlink work. Pick by the job you need done, then compare providers on price and the fields returned.

### How many tools should an AI agent have?

Only as many as the job needs. Each tool costs context, and too many make the model choose worse. A gateway can show a small set of tools that search and run a larger catalog.

### Do AI agent tools need APIs?

Most do. A tool is usually a thin definition over an API call. Some tools run locally, such as reading a file, and need no API.

### How do I give an agent access to data sources?

Connect it to a gateway or an MCP server that fronts the data providers. With looot the agent connects once and finds endpoints by describing the job.

### Are AI agent tools free?

Some are, with limits. Data and tool APIs that cost real money to run charge per call or per plan. With looot, searching and inspecting are free and each run costs its listed price.

## For agents

### Map my agent's tools

```text
Use looot to show me what tools I can give an agent.
1. Set up https://looot.ai/skill.md.
2. List the tools you received and what each one does.
3. For each of these jobs, use discover and tell me the cheapest endpoint and its price: web search, page as markdown, work email lookup, keyword volume.
4. Do not run any endpoint. Tell me my balance.
```