Voice agents you can prove: what Pinecall's launch gets right about testing and guardrails
Pinecall (the open-source one at pinecall.io), hands-on notes. Independent; not affiliated with Pinecall. Links to my free cheat sheet and starter pack are at the end. Most voice-agent demos share a failure mode: they w
Pinecall (the open-source one at pinecall.io), hands-on notes. Independent; not affiliated with Pinecall. Links to my free cheat sheet and starter pack are at the end.
Most voice-agent demos share a failure mode: they work beautifully on the happy path, in a quiet room, with a cooperative caller — and nobody can tell you what happens on the fourth call, with a noisy line, an angry caller, and a booking that shouldn't have happened. Pinecall, which open-sourced its whole platform (Apache-2.0, built on LiveKit) for its Product Hunt launch on November 16, is interesting because it treats that gap as the product.
The agent is the small part
The founder's launch post puts it well: with every client, the agent was the small part; the work was behind it — who answers which number, when a call passes to a person, what happens after it hangs up. So Pinecall is a platform around the call:
- An agent is one class (TypeScript, Python, or Ruby): fields are its state, tools are what the model may do, docstrings are the prompt. Or no code at all — a form in Studio — or your coding assistant driving it over MCP. Same runtime either way.
- The same agent answers the phone, a website, and WhatsApp (in writing).
- A supervisor can listen in, whisper to the agent, or take the line and hand it back. "A person one tap away" is a feature, not a slogan.
- You can run it on their cloud — or self-host it free, with every feature, down to a one-command local runtime with open models (Whisper large-v3-turbo, Gemma 4 or Nemotron on Ollama, Kokoro for the voice) on your own GPU, with no vendor bill at all.
Proof, in four layers
- Unit tests — the class as software, no network, no model.
-
Goldens — written conversations with what must be true at the end, replayed through the agent (
pinecall test) under every model you name. In CI, goldens are the gate before a deploy. -
Simulated callers (personas) — a model plays the caller, in writing or out loud, even over noise and packet loss (
pinecall simulate). The platform can even generate tough callers from what your agent is. -
Judges — every finished call gets a verdict at hang-up. Three of the built-in judges are the ones I'd switch on first:
-
disclosed— the caller actually learned they were talking to an automated assistant (and the call fails if the agent claimed to be a person); -
consent— an irreversible tool (a booking, a cancellation, a charge) ran only after the caller agreed to that specific action; -
honoured-stop— a caller who asked not to be called again was put on the do-not-call list during that call.
-
A broken call becomes a case you keep — so a fixed failure is regression-tested forever. That's the part most platforms leave as an exercise for the reader.
Guardrails that refuse, not just advise
Outbound calling is where voice AI gets legally and ethically hot, and Pinecall's answer is a rule chain every dial passes, in order: the number must be valid E.164 (shape); it must have contacted your org before (stranger — cold lists are refused by design); it must not be on your do-not-call list (do_not_call — and a consent sent with a dial never lifts that); a +1 number needs consent on file (no_consent); and it must be inside allowed hours in the number's own time zone (quiet_hours — a +1 number never rings before 8:00 or after 21:00). Every call also opens with a spoken disclosure — "This is an automated assistant calling on behalf of …" — plus a recording notice when recording, logged as the call's first turn so you can prove what the caller heard.
Used the way it's designed — your own business's reception, and callbacks to people who already know you — this is what responsible voice automation looks like. Used for bulk cold outreach, it mostly refuses. Both are selling points.
The honest pricing picture (checked November 18, 2026)
Self-hosted is $0 forever, every feature. On Pinecall Cloud: a free sandbox; Starter $29/mo (500 minutes on their models included, then $0.06/min — $0.02 on your own keys), Team $99/mo (2,000 minutes, then $0.05 / $0.018), Scale $499/mo (10,000 minutes, then $0.045 / $0.015). Hosted apps are $5/month each. Every call's record carries its per-provider cost, line by line.
One warning that tells you a lot about the state of a launch-week product: the official pricing page says the free sandbox includes 120 minutes and 30 simulations, while two official docs pages still describe the free trial as 30 minutes and 10 simulations. Both are official; they can't both be your limit. The fix is one command — pinecall limits in your own org — and it's the number to plan around, not either page. (The founder is equally plain elsewhere: the platform "has run in production for a handful of companies, not thousands, so there are rough edges.")
Try it in 10 minutes (free)
npm i -g pinecall # CLI, Node 24+
pinecall login
pinecall new acme --agent front-desk
cd acme && npm install && pinecall start # pinecall chat, in a second terminal, talks to it
No vendor keys needed to start: Pinecall's keys run Deepgram's ears/voices and Gemini's flash-lite; anything else (Claude, GPT, Cartesia, ElevenLabs) runs on your own key, added via stdin and stored sealed. Start in the sandbox, write the goldens first, hear it from your own phone (pinecall line from +…) before any customer does.
Free: I condensed the verified pricing, the 4-command start, the front-desk loop, and the 5 dialling rules into a one-page cheat sheet: https://payhip.com/b/0pifs
Full starter pack: a 21-section tutorial, deployment templates with go-live gates, 3 complete reception workflows (front desk, consented callbacks, after-hours), and 6 copy-paste assistant prompts: https://payhip.com/b/bdEXL
Disclosure: the pack is my own independent product, sold on Payhip. The cheat sheet is free and covers the essentials. No answer-rate or conversion promises anywhere — just the setup, and the tests to prove it works.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.