> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.jambonz.org/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.jambonz.org/_mcp/server.

# Agent Studio

Agent Studio is where you build a voice agent, call it from your browser, publish it to a phone number, and see every
call it takes. It connects to your jambonz account (jambonz.cloud or your own jambonz) with an API key, and uses the
carriers, phone numbers and speech and LLM credentials you already have there.

## Four kinds of agent

You choose the kind when you create an agent. It decides where the agent's thinking happens; everything else (phone
numbers, transfers, tools, recording, call history) works the same for all four.

| Kind                        | What it is                                                                                                           | Choose it when                                                                                                                                      |
| --------------------------- | -------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Cascaded voice pipeline** | jambonz runs speech recognition, a language model and a voice that you choose                                        | you want control over each part: any LLM, the widest choice of voices, latency you can tune per stage                                               |
| **Realtime voice API**      | one vendor runs the whole voice loop: OpenAI Realtime, OpenAI GPT-Live, Deepgram Voice Agent, Rime Prism or Ultravox | you want the lowest latency and the most natural turn-taking, with one vendor to configure                                                          |
| **Connected agent**         | an agent built on another platform, Dialogflow CES or ElevenLabs Agents, put on the phone by Studio                  | the agent already exists there, or your team builds agents there                                                                                    |
| **Custom harness**          | your own code decides what to say; jambonz listens, takes turns and speaks                                           | the conversation logic is yours to own: your own models, retrieval, routing and checks. See [Custom harness](/agent-studio/custom-harness/overview) |

## What every agent gets

* **Phone numbers**: route your jambonz numbers to an agent, for inbound calls, or place outbound calls from it.
* **Transfers and hangup**: hand the caller to a person or another system (blind or warm, with a spoken brief), take
  them through a secure detour (for example, a payment) and back, and end calls cleanly.
* **Tools**: HTTP tools and MCP servers the agent can call during a conversation.
* **Testing**: call the agent from your browser or your phone and watch each turn as it happens.
* **Versions**: publish a draft when it is ready; roll back to an earlier version at any time.
* **Call history**: recordings, transcripts and each turn's timing and tool calls for every call.

## Next

* [Getting started](/agent-studio/getting-started): connect your jambonz account, create an agent and call it
* [Building an agent](/agent-studio/building/prompt): prompt, tools, knowledge, conversation, transfers
* [Testing](/agent-studio/testing) and [Publishing](/agent-studio/publishing): try it, then put it on a number
* [Calls and usage](/agent-studio/calls): every call's turns, timing and recordings
* [Custom harness](/agent-studio/custom-harness/overview): build an agent whose thinking runs in your own code