Launched this week

Pinecall
The open-source platform for voice AI agents
103 followers
The open-source platform for voice AI agents
103 followers
Voice agents you can prove. Pinecall is the open-source platform for voice apps: build in a form, in code, or from your AI assistant over MCP. Test with simulated callers, get every call logged and judged, and keep a person one tap away. On our cloud or your own Kubernetes. Apache-2.0, built on LiveKit.







Free Options
Launch Team / Built With

AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted

Hi Product Hunt, I'm Bernardo, one of the founders of Pinecall.
Our small team has built voice agents for a handful of companies, and we never wrote a word about it in public. Today all of it is open source, under Apache-2.0.
We spent a long time on the wrong layer: our own integration with every transcriber, model and voice. LiveKit already did that better, so we threw ours away and built on it. What was left is everything around the call, and that is Pinecall.
A framework for voice apps, not only agents. With every client the agent was the small part; the work was behind it: who answers which number, when a call passes on, what happens after it hangs up. Write it in TypeScript, Python or Ruby, build an agent in a form in Studio, or let your coding assistant do it over MCP. They all run on the same platform.
Tests before anyone calls. Written conversations, and simulated callers on a real spoken line, over noise and packet loss.
Every call judged. A verdict at hang-up, a log you read entry by entry, and a broken call becomes a test case. A supervisor can listen in, whisper to the agent, or take the line and hand it back, from the console or a phone.
Built for the bad day. A worker that dies mid-call hands the call to another worker with its history. A deploy drains live calls instead of cutting them. Speech-to-text and voice fall back to the next vendor.
Run it yourself. Terraform for the cluster and Helm charts that scale workers on live calls, down to zero. The same charts our own cloud runs on.
Or run it with no vendor at all. One command brings the whole runtime up on your machine with open models: Whisper large-v3-turbo, Gemma 4 or NVIDIA Nemotron on Ollama, and Kokoro for the voice. On an NVIDIA card or Apple silicon, no account and no vendor bill.
Try it now: talk to the agent on pinecall.io, or build your own in the sandbox with 120 free minutes and no card.
It has run in production for a handful of companies, not thousands, so there are rough edges. If you have put a voice agent in front of real callers, I'd like to hear what broke first.
the supervisor whisper is a neat touch. when someone takes the line, does the caller hear a handover or is it seamless?
Thanks @neerajnathany! It's your call, there's no forced handover sound.
A whisper only reaches the agent, the caller never hears it. You can also type a line and the agent says it verbatim in its own voice, so the caller notices nothing. When you take the line, the agent goes quiet and the caller hears you. If you want it smooth, have the agent say something like "one moment, I'm putting a colleague on" right before you take it, then give the line back when you're done.
And when the agent itself asks for a person, the caller waits on hold music until a supervisor picks up.
@bernatch the line you type that the agent says in its own voice is a clever trick, the caller would never guess someone else is steering
@neerajnathany exactly! The idea is to give the supervisor full control without breaking the conversation flow. The agent becomes a kind of invisible interface between the caller and the human behind it.
The testing story is what stood out to me — simulated callers that interrupt and talk fast, and a broken call becoming a test case, is exactly the part most voice agent demos skip. Real callers don't wait politely, and I like that every call gets logged and judged instead of just hoping it went well. Being open-source and runnable on my own infra with open models is the other big draw for me. Going to try the sandbox with the 120 free minutes.
Thanks @jcbuilds! That's exactly the kind of feedback we were hoping to get. Let me know how you find the sandbox!
simulated callers who interrupt and talk fast before real calls, smart. congrats Bernardo!
Thanks @mohitvishwakarma! That's exactly the idea. Real callers don't wait for the agent to finish speaking, so we wanted to make sure you can test those situations before going live. Appreciate the support!
@bernatch makes sense, interruptions are where most voice agents break. good luck with it
The simulated callers piece is what caught my eye, testing voice agents by hand is painful. Are the simulated callers scripted scenarios or do they improvise, and can the judge flag when the agent drifts off the intended flow?
Thanks @galdayan! Both, and they're two different tools.
Goldens are the scripted ones: a conversation written down, with what the caller says and what must be true at the end. They're deterministic and run in CI before every deploy.
Personas improvise: a model plays a caller with a goal, a manner and the facts they know (someone in a hurry, a wrong number, someone who keeps interrupting), in writing or out loud on a real line. They find the things you didn't think to script.
On drift: every call is judged at hang-up and each verdict comes with its reason. A persona can carry what it expects from the agent and a judge holds the call to that, and you can write your own judges as a plain question about the call. A real call that breaks a judge is kept as a case you can reproduce, fix and add to the suite.