Your Own Cartesia Key Doesn't Make Your Voice Agent Self-Hosted

Putting your own Cartesia key into a voice agent feels like taking control. What it changes is whose name sits on the contract. Every line the agent speaks is still voiced on Cartesia's servers, and every word the caller says is still transcribed there.

That gap is about to matter to a lot of teams. Cartesia stops hosting Line SDK agents on December 1, 2026. Line code you run yourself keeps working, so plenty of people will move their agents onto their own machines and call the job done. A self-run Line agent still connects to Cartesia through a self-hosted deployment URL, though. Your code moved. The call didn't.

So before anyone types "self-hosted" into a security questionnaire, it's worth being exact about what Cartesia opens up and which pieces you'd have to replace yourself.

Cartesia is more open than people think, in the wrong places

Credit where it's due. The Python and JavaScript SDKs are Apache-2.0, and so is the Line SDK for code-first agents. That's real open code you can fork and read.

The parts that actually answer the phone are closed. Managed Agents is Cartesia's product running in Cartesia's cloud, and the Sonic and Ink model weights aren't published. On Enterprise, the sales team can arrange an on-premise or private-cloud deployment of the models. That's a licensed private deployment under a contract: it runs on your hardware, and you still can't read or change the code.

I find it easiest to treat this as a question of layers, because "open-source alternative to Cartesia" means something different depending on which layer you want to own.

Owning the agent code is the easy one. Keep running Line on your own servers, or move to LiveKit Agents or Pipecat. The catch is everything around the agent. Telephony, call history, QA and a dashboard all turn into things your developers build and maintain.

Owning the whole platform means replacing Managed Agents as a product, with the builder and the phone side included. That's the layer we work on. Dograh, the open-source AI voice agent platform we build, is BSD-2-Clause and covers it, and you can self-host it or run it on Dograh Cloud.

Owning the voice is the hard one, so it gets its own section.

Open voices are usable, and they're behind

Be honest with your stakeholders here. On the Artificial Analysis Speech Arena, where listeners vote blind, Sonic 3.6 sits in fourth place. Kokoro 82M is 62nd and Chatterbox is 80th. Swapping Sonic for an open model is a real quality trade, and you only learn how big it is for your callers by testing on actual phone calls.

If you still want to go open, Kokoro 82M (Apache-2.0) is the quick start. It's small and fast, with fixed voices and no cloning from a reference clip. Kokoro is also the default model in our Local Models option, which runs through Speaches, so there's no extra wiring.

Chatterbox (MIT) is the one to test if you need cloning, since it copies a voice from a short clip. Every output carries Resemble AI's audio watermark. One practical snag: the common Chatterbox servers return WAV or MP3 files, while our local voice slot expects raw 24 kHz audio, so plan for a server that returns that format.

On the speech-to-text side, Whisper models run through the same Speaches connection, with faster-distil-whisper-small.en as the default.

Read licences carefully past those two. Fish Speech ships under the Fish Audio Community License, which isn't OSI-approved and needs a separate commercial licence for public services. F5-TTS and XTTS-v2 are licensed for non-commercial use only. Downloadable weights don't automatically mean you can put them in a product.

You don't have to drop Sonic to own the platform

Here's the middle path a lot of teams miss. Cartesia is a built-in provider in our platform, so you can keep the voice your callers already know. You pick Sonic 3.6, 3.5 or 3, paste any Cartesia voice ID (cloned voices included) and add your own key. For speech-to-text, Ink 2 covers English and ink-whisper handles other languages.

That gets you an open platform with a closed voice. It's a perfectly good place to be, as long as everyone understands the audio still goes to Cartesia under your key.

If calls must never leave your servers, the models have to live there too. Point speech-to-text and the voice at Speaches, add an open-weight language model, and turn off the anonymous usage telemetry a self-hosted install sends by default by setting ENABLE_TELEMETRY=false and removing POSTHOG_API_KEY. A self-hosted install runs 10 calls at once out of the box, and your servers and phone carrier set the real ceiling.

That's the actual trade. The vendor invoice shrinks, and the GPUs and upgrades become your team's job. For a small team at low volume, a hosted platform can honestly be the better use of time.

If you've already moved your Line agents off Cartesia's hosting, did the call path change as much as you expected, or is Cartesia still in every call?

The layer-by-layer table and the full self-hosting setup are in our breakdown of open-source replacements for Cartesia's platform and voices.