Voice agents on your own GPUs, with no audio sent to AI model providers
Dograh is the self-hostable voice agent platform with an open-source workflow builder. Put open-weight speech and language models on a GPU you own while Dograh runs the call around them, so no audio goes out to a model provider.
Two layers on one machine
Fill the model layer yourself
Open weightsmodels you download and run
Each model in the call can be an open-weight release you host, so no provider ever hears the audio.
→ Read each licence, some ban commercial use
Size it to one card
Quantizationsmaller weights, a little less accuracy
A 4-bit 27B language model fits beside a small speech-to-text model on a single 24GB GPU.
→ Check the speech model's memory before adding it
Keep the orchestrator as it is
Orchestrationthe call flow, kept apart from the models
Dograh runs the call flow and telephony whichever models sit underneath, so moving them local later means changing settings.
→ Start on hosted models, move local when ready
Choose what fills the model layer
Only this column keeps the call audio on your hardware.
Quick to start, but the provider still receives the audio.
None of this changes when the models move.
Check these before the first real call
- Licences readEach model allows commercial use, and attribution is in place where the licence asks for it.
- Memory measuredBoth models loaded together at your context length, with room left on the card.
- Phone call testedLatency was measured on a real phone call, since the phone network adds delay a browser test hides.
- Voice choice checkedIf the design needs a cloned voice, the TTS model can clone. Magpie cannot.
- No outside endpointsEvery model URL in the agent points at your own hardware.
What owning the model layer buys you
Audio stays on your hardware
Self-host Dograh with open weights beside it, and no part of the call reaches a model provider.
No hop between stages
Colocated models pass audio to each other on one machine, so there is no round trip to wait on.
Flat cost at any volume
A GPU you own costs the same at 200 calls a day or 20,000. Engineer time still costs money.
Move local when you're ready
Start on hosted models and swap the model layer later without rebuilding the agent.
What changes when you own both layers
| What you get | Dograh | Hosted voice platforms |
|---|---|---|
| Open-source orchestrator you can read and change (BSD-2) | ✓ | ✕ |
| Runs on your own servers | ✓ | ✕ |
| Open-weight models on your own GPU | ✓ | ✕ |
| No per-minute platform fee when self-hosted | ✓ | ✕ |
Build the local stack, step by step
Does bring your own key give me data sovereignty?+×
No. It changes who holds the billing relationship. The audio still travels to the provider and is processed on their servers.
Is Dograh a voice model?+×
No. Dograh is the orchestrator that runs the call flow and telephony. The speech and language models are a separate layer you choose.
What hardware do I need to run open weights?+×
One 24GB GPU fits a 4-bit 27B language model beside a small speech model such as Parakeet TDT 0.6B v2. Benchmark your own box.
Will self-hosting make my agent slower?+×
Running every stage on one machine removes the network hop between them. Test on a real phone call, because the phone network adds its own delay.
Can I start hosted and move to open weights later?+×
Yes. The orchestration layer stays the same, so moving the models onto your own hardware is a configuration change.
Are open-weight models the same as open source?+×
No. Open weights means you can download and run them. Licences vary, and some forbid commercial use, so read each one.
Keep the orchestrator. Change the models later.
Self-host in one command, or start on Dograh Cloud.
$ curl … start_docker.sh && ./start_docker.sh