Introducing Tone: the open-source voice AI platform.
Why we're building Tone in public, the gaps in the closed-source incumbents, and what shipping looks like from here.
Voice AI is having its moment. Synthflow, Vapi, and Retell have shown what's possible when you wire a competent LLM to a phone line. But the moment you try to run one of these stacks in production — for healthcare, for finance, for anything where the answer to "where does my data go" matters — the same questions keep coming up.
The gap we kept hitting
Every team we talked to before starting Tone had the same shortlist:
- "Can I see the prompt the agent is actually running?"
- "Can I run this on my own GPUs?"
- "Can I swap the LLM, the voice, or the telephony provider without rewriting everything?"
- "Can I read the source when something breaks at 2 a.m.?"
Today the answer to all four is no. Tone changes that.
What ships in v0.1
The first cut is opinionated and small on purpose. You get:
- A workflow runtime that compiles to a deterministic state machine.
- An eval harness with persona-based regression tests built in.
- A telephony adapter that speaks Twilio, Vonage, and SIP out of the box.
- A self-host story that's
docker compose upfor development and a Helm chart for production.
git clone https://github.com/tonehq/tone
cd tone && docker compose up -dThat's it. Point a Twilio number at the resulting endpoint and you've got an agent answering the phone.
Why open source, and why now
The closed-source voice incumbents are great products. We're not trying to replicate them — we're trying to build the thing they look like when the lock is off. If you've built on top of a hosted-only platform before, you know the feeling: every decision they make is a decision you can't unmake.
We think voice AI is too important an interface to live behind a vendor login. The code is MIT. The roadmap is on GitHub. The RFCs are in the open.
If that resonates, star the repo, jump into Discord, or open a PR. Field notes from here on in.