Tool actions
Let the agent do the work. Not just describe it.
Give your voice agent scoped access to your own APIs. During the call it looks up live information, takes the actions you approved, and tells the caller the result in plain words while they are still on the line.
Answers help. Finished actions help more.
The phone call becomes a way into the systems your business already runs on. Your systems stay the source of truth: the agent reads from them and writes to them through your own endpoints.
Example call, with sample data
What a tool can do. Read, act, stay in your stack.
Use current data
Availability, order status, customer records, prices, balances: any fact your API exposes to the tool.
Take bounded actions
Open a ticket, update a record or start a workflow in your own API, with typed arguments your endpoint validates.
Keep your architecture
Tools are defined on the voice gateway, which calls your HTTPS endpoints directly. No middleware to run.
From conversation to API. A controlled path.
A tool definition tells the model when to use it and what to collect, and tells the gateway what request to make and how to turn the answer into a sentence.
- 1
Define the tool
A name, a description of when to use it, and a strict schema for its arguments.
- 2
Keep secrets out
The tool names an environment variable on the gateway. The key itself never goes in the tool file or the prompt.
- 3
Validate at your endpoint
The gateway checks arguments against your schema first. Your endpoint still authenticates the request and rejects anything outside your policy.
- 4
Say only what matters
A response template turns your JSON into one speakable sentence, so the agent never reads field names aloud.
Least privilege. Checked by the gateway.
A tool is not a blanket permission. Each one has its own method, endpoint, schema, timeout, credential and scope.
Custom tools reference- Owner-only toolsA tool marked private is left out of public calls entirely, and refused if it is called on one anyway.
- HTTPS to public hosts onlyPlain HTTP and private, loopback or cloud metadata addresses are refused, when the file loads and again after the arguments are filled in.
- Checked before useReserved names, unsafe URLs and malformed schemas fail validation. A file that fails loads none of its tools, and the agent keeps running without them.
- Fast and directHTTP tools run straight from the gateway with a short timeout, without a round trip through a slower general-purpose agent.
- Bounded repliesYour endpoint must answer with JSON and a success status, and what reaches the model is capped in size.
What one tool can unlock.
Live inventory
Is it in stock, and where.
Order and account status
Read live from the system that owns it.
Quotes
A price calculated by your own system.
Tickets
Opened with the caller's own description of the problem.
Return requests
Started in your system, by your rules.
Your own workflow
Anything your API can start.
Questions.
Can a tool call any API?
Any HTTPS endpoint on a public host that passes the gateway's checks. Your endpoint stays responsible for authentication, authorization and business rules.
Who sets up the tools?
On an agent we run for you, our team builds and maintains the tools with you during the build. On a gateway installed on your own server, you can edit the tool file yourself.
Where are API credentials kept?
In the gateway's environment file. The tool definition names the variable and never holds the token. Whoever controls the gateway's server can read that file, so use a key scoped to what the tool needs.
Can public callers use owner tools?
Not when the tool is marked private. It is left out of the agent's tools on public calls, which is stronger than a prompt instruction.
Does the agent confirm before it changes something?
A tool that changes something should say so in its description, for example "use only after the caller confirms", so the agent asks first.
How many tools can one agent have?
Up to 20 custom tools, including the ones connected apps such as Shopify or Wix add. Fewer, sharper tools work better, because every tool adds to each turn's prompt.
Which call should your agent take first?
Tell us which requests your team handles most often. We'll suggest a practical place to start.