> ## Documentation Index
> Fetch the complete documentation index at: https://aipro.placetel.de/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Build a voice agent

> Describe the agent, review the instructions, tune how it talks.

Your token needs `voice_agents:read` and `voice_agents:write`. You never call
MCP tools yourself: you describe what you want, the assistant picks the tool,
and you review before it runs. Tool names below are there so you can recognise
them in your client.

## 1. Describe the agent

Say who calls, what the agent should handle, what it must hand over, and how
it should sound:

> Create a voice agent "Empfang" for our car workshop. Callers book service
> appointments or ask about opening hours. Anything about invoices goes to a
> human. Formal address, friendly and brief.

The assistant checks existing agents (`list_voice_agents`) and creates the new
one (`deploy_voice_agent`). Despite the name, this does not connect a phone
number; that happens in the dashboard later. Expect it to report the new
agent's ID.

## 2. Review the instructions

The instructions are the agent's prompt. The assistant drafts them following
the [Prompting Guide](https://aipro.placetel.de/docs/en/ai-prompts.md) and should show them before saving
(`set_agent_prompt`). Check:

* Identity and tone match what you asked for.
* Each task says what information to collect.
* Hand-over and "never do" rules are explicit.
* Spoken-language details are right: numbers, dates, times as a person says them.

Say what to change; the assistant updates the draft and saves it again.

## 3. Tune how it talks

Settings come in three groups: `prompts` (greeting and behaviour),
`voice_config` (voice, language, speech) and `knowledge_base`. Ask in plain
words:

> Greet with "Guten Tag, Autohaus Berger, wie kann ich helfen?", speak a bit
> slower, and use a female voice.

The assistant reads the settings schema first (`get_voice_agent_concern_schema`),
then changes **one setting per call** (`update_voice_agent_concern`) and
reports each result. Two things to watch in its answers:

* For choice fields it must send the option's exact value from the schema,
  not the label you see in the dashboard.
* Some changes have side effects: switching the conversation mode resets
  dependent settings. The response lists them under `effects`; the assistant
  should mention them and re-read the schema before continuing.

## 4. Finish

Connect a [phone number](https://aipro.placetel.de/docs/en/ai-numbers.md) in the dashboard, then
[monitor the first calls](https://aipro.placetel.de/docs/en/mcp/monitoring.md). Later changes work the same way: ask,
review, confirm. Renaming uses `update_voice_agent`; `delete_voice_agent`
removes an agent and should only run after you confirmed it explicitly.

Next: [add tools](https://aipro.placetel.de/docs/en/mcp/tools.md) so the agent can look things up or book.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.