# How can my AI agent find the right tool at runtime?

Use Agentik discovery to find tools by task, then inspect exact contracts.

By [Agentik](/about.md). Updated: 2026-10-19.

Agentik lets an agent discover candidate tools for a task and inspect the chosen operation before execution. This helps the agent select a relevant capability without treating a long catalog as an undifferentiated instruction block. Discovery identifies possibilities; inspection establishes the schema, access requirements and price the application must respect.

- [Connect your agent](/docs/clients.md)

## Describe the required operation and output

A useful discovery request names the task, source and expected output: find public company pages, verify a supplied email, or generate a square product concept. A vague request for “marketing tools” produces a less actionable shortlist.

After discovery, inspect the exact candidate rather than rely on its name. Confirm required inputs, supported account mode, result format, pricing unit and asynchronous behavior. A provider listing does not imply that all of its endpoints are served.

## Validate the selected contract before dispatch

Public discovery can be read-only. Actual execution, including a zero-price provider operation, requires authority for an Agentik workspace. Finding a tool therefore does not establish permission to run it.

Keep the inspected contract with the planned input and budget decision. If an implementation is unavailable or the schema no longer fits, discover another candidate and inspect it independently. Do not silently substitute a provider while keeping the previous price or input assumptions.

## Keep the agent’s working context focused

Retain the small set of selected tool contracts needed for the current job. Your application can choose how to cache or refresh them; Agentik discovery does not itself prove lower model token usage or a benchmark improvement.

At the final planning checkpoint, show the user the required inputs, ordered calls and scope. Execution results remain data to validate, and the agent should preserve run identity when following pending work.

## What you pay for

Supported paid tools use Agentik's prepaid workspace balance. Inspect the selected tool's pricing units and maximum before approving a run; a reservation can differ from its settled charge. Funding fees, taxes and your model's usage are separate. Current terms are on the Pricing and Billing pages linked below.

## Discover a verification capability precisely

A supplied email requires verification rather than contact discovery. The flow below is a selection plan, not a provider run.

Discover a verification capability precisely

| Step | Check |
| --- | --- |
| Describe | Verify one supplied address and retain the exact verdict. |
| Discover | Find supported email-verification candidates. |
| Inspect | Read required email input, result states and pricing. |
| Authorize | Confirm workspace authority and approved spend before execution. |

## A prompt to try

[Connect your agent](/docs/clients.md), then use this prompt with your connected assistant. Keep credentials in the authorized connection.

```text
Find a tool for [specific operation] that returns [required output] from [source]. Use Agentik discovery, then inspect_tool for the most relevant candidates. Check required inputs, availability, access mode, output, units and run handling.

Explain why the selected operation fits and what remains unknown. Do not assume discovery grants execution permission. Show the proposed input and maximum, then wait for approval before paid calls.
```

## A related workflow

[Account research](/account-research.md). Their site, latest news, official filings and the profiles of the people you’ll meet, in a one-page brief with sources.

Example tool-call maximum: at most $0.12.

One US-listed company, 3 pages of its site, 20 news results, its filings, its LinkedIn page and 3 attendees.

This catalog-based reservation maximum belongs to the playbook’s inputs, separately from this guide’s worked example. It is not an executed charge. Optional calls, funding fees, taxes and model usage are excluded.

- [firecrawl/scrape](/providers/firecrawl/scrape.md)
- [serper/news](/providers/serper/news.md)
- [sec-edgar/company-filings](/providers/sec-edgar/company-filings.md)

[View every step, input and tool in the playbook](/account-research.md).

## Questions

### Can a public discovery connection execute tools?

No. Discovery and inspection are read-only. Execution requires authorized workspace access, including for a tool priced at zero.

### Can the agent substitute another provider automatically?

It should inspect the replacement contract and respect the approved task and budget. Availability alone does not make schemas or prices interchangeable.

## Sources and next steps

Agentik publishes this guide. The editorial update date is 2026-10-19; provider terms and availability can change.

- [Discovery contract](/docs/discovery.md)
- [Authentication](/docs/authentication.md)

### Continue with Agentik

- [Discovery documentation](/docs/discovery.md)
- [REST quickstart](/docs/quickstart.md)

## Related guides

- [MCP or REST for tool access](/ai-agent-mcp-vs-rest.md)
- [One API for your agent's external tools](/one-api-for-ai-agent-tools.md)
- [Error handling for paid tool calls](/ai-agent-tool-error-handling.md)

[Browse all agent tool guides](/agent-tool-guides.md)

---

[HTML page](/ai-agent-tool-discovery) · [Agent index](/llms.txt)