Everything your AI agent needs to talk.

1. Say what it should doWrite the instructions as you would brief a person: who it is, what it helps with, how it should speak.
2. Give it a voiceChoose a language and a voice, and listen before you pick. Set the pace and the mood.
3. Talk to itPress Test agent and have the conversation your callers will have. Change a line, try again.
4. Put it where people areOne line of HTML on your website, or one request from your server to start a call in your app.

Pick a voice engine, a language and a voice, then set how fast it talks and the mood it speaks in. Listen to a voice’s sample before you choose it.

Three engines are on offer. Cartesia Sonic speaks every language on the list. Sarvam Bulbul has voices recorded in Hindi, English and seven more Indian languages, in speaking styles from customer care to news reading. Rumik Mulberry styles a Hindi, Hinglish or English voice from a one-sentence description.

• Each reply is written in the language’s own script, then spoken: Hindi in देवनागरी, Tamil in தமிழ்.
• Hinglish is a language of its own here, for callers who move between Hindi and English in one sentence.
• An agent can move between Hindi and English when the caller does, and a single call can be started in another language.
• Every combination of voice, language and model costs the same per minute.

Timing is what makes a voice agent easy or tiring to talk to. Crackle handles it for every agent, with nothing to configure.

• Start talking and the agent stops mid-word and listens, then picks the thread back up.
• A short “okay” or “mm-hm” while it talks is taken as a listener nodding along, not an interruption.
• It opens the call your way: a fixed line, a line written for this caller, or silence until the caller speaks.
• On voices that can perform them, a written laugh or sigh is acted out, never read aloud.
• When the caller is done it says one short goodbye and hangs up.

A tool is one of your endpoints, described in plain words. The agent decides when to call it, waits for the answer and tells the caller what happened.

• During-call tools run when the caller asks for something; before-call tools fetch what the agent should know before it says hello.
• Try a tool from the studio before a caller ever reaches it.
• Built-in tools let the agent end the call or switch language on its own.

Give an agent your documents and it answers callers from them. A short set goes straight into its instructions; a longer one is searched when a caller asks, with a short holding line while it looks.

• PDF, Word (.docx) and text files up to 25 MB each, or pasted text, up to 500 documents a workspace.
• When a lookup finds nothing, the agent says it does not have that information. It is told never to guess a fact about your company.
• Try a question first and see exactly what the agent would be handed.
• Files are encrypted before they are stored, and an agent searches only the documents switched on for it.

Write blanks such as {{customer_name}} into the instructions, the greeting or a tool, and fill them in when each call starts. The same agent greets every caller by name.

• Each variable can have a default for when you have no value.
• A before-call tool can look the caller up in your own system instead.
• Your own data, such as an order ID, can ride along with a call and come back in every webhook for it.

Switch on the call button for an agent and paste the line the studio gives you. A visitor presses it, allows the microphone and is talking to your agent.

• The line carries a publishable key that can only start calls with that one agent.
• You choose which websites may use it.
• Prefer your own button? A small JavaScript API starts and ends calls and hands you live captions.

Your server starts a call with one request and hands your app a join token. Calls run on LiveKit, so any of its client SDKs will do.

• Your API key stays on your server. The app only ever holds a token for one room.
• Live captions, and whether the agent is listening, thinking or speaking, arrive as the call goes.
• A “Didn’t catch that” button can ask the agent to say it again.

Every call is kept with its transcript, how long it ran, what it cost and why it ended. Webhooks bring the same to your own server as it happens.

• Events for a call starting, the caller joining, each line spoken and the call ending.
• Each delivery is signed, so you can check it came from Crackle.
• Your server can have the agent say a line in a live call, or end it.
• Calls are billed by the second, and a call nobody joins costs nothing.

The studio is where an agent is made: instructions, greeting, voice, tools and documents on one page, and a Test agent button that puts you on a call with it. The API works on the same agents, so nothing is rebuilt when you wire it into your product.

• An agent made in the studio takes calls started through the API, and the other way round.
• Workspace API keys you can create and revoke, and teammates with their own roles.
• Keep an agent in your repository as a file and sync it with a script.