Dynamic prompting with full control, in plain code you own. Your coding agent can build, test, and simulate conversations before it ships.
npm create curb.js .
Your prompt is React, so it follows your state. At every moment of the call, the model gets only the instruction and tools that fit.
CurbJS
export default function ReportClaim() {
const [verified, setVerified] = useState(false);
if (!verified) return <>
<Instruction>Ask for the policy number.</Instruction>
<Tool name="verify_policy" …
onSubmit={() => { setVerified(true); … }} />
</>;
return <>
<Instruction>Ask what happened.</Instruction>
<Tool name="open_claim" … />
</>;
}Model context
Instruction
Ask for the policy number.
Ask what happened.
Tools
verify_policy
open_claim
A tool your code does not render is not in the request, so the model cannot call it.
Each job gets its own instructions and tools, and the model opens it when the call needs it.
Speech to text, text to speech, turn-taking and interruptions are built in.
The agent asks first, and only the caller’s yes runs the tool.
The model gets one step at a time, and the last tool appears only at the end.
Transfer to a person, with a fallback when nobody answers.
Switch voice, language or model during a call.
Keep a whole call as a test and run it from the terminal.
Deploy to KugelAudio, or run the same project on your own servers.
Write a call down once and run it after every change, with no microphone. When a real call goes wrong, your coding agent finds the cause, fixes it and keeps the call as a test.
$ npx curb test
--- opens a claim for a burst pipe
caller: A pipe burst at home last night
agent: Your claim is open.
pass called open_claim
--- asks what happened before opening a claim
caller: I need to report some damage
agent: What happened, and when?
pass did not call open_claim
--- answers a question without opening a claim
caller: Is water damage covered?
agent: Water damage is covered.
pass did not call open_claim
3 of 3 scenario(s) passed in 0.6s.curb test run on the quickstart agent, with a test model standing in for the LLM.Every part of the pipeline is a prop. Run it all on KugelAudio with one sign-in, or swap one part at a time.
npm create curb.js .In an empty folder. Keep the defaults and the agent runs on KugelAudio with one sign-in.
npx curb loginOpens your browser once. There is no key to copy.
npm run devType to your agent, or start a call, and watch what the model has.
npx curb testRuns whole conversations from the terminal, with no microphone.
Put the caller through to a person, with a callback when nobody answers.
Hand over by code, to a skill with its own voice, or to a person with a note.
Who speaks first, the order of a conversation, what needs a yes.
Switch language, voice or model during a call.