Run it yourself.
Clone the repo, docker compose up, you're running. Forever free, BSD 2-Clause, your servers, your data.
Fork on GitHubTHE SELF HOSTABLE VOICE AI FULL STACKOSS ALTERNATIVE TO VAPI & RETELL
[On Device] voice agents_
Open-source workflow builder for voice agents. Fork every box, bring your own models and run the whole stack inside your own perimeter.
#1 Product of the Day #1 Product of the Week NVIDIA Inception
LIVE · YOUR STACK
production.agent
$ dograh up --stack self-hosted
✓ asr cartesia ready · 84ms
✓ llm dograh ready · 212ms
✓ tts deepgram ready · 96ms
● call connected · audio-to-audio
1 hopaudio-to-audio
70+languages
0vendor calls
01 / 02 · KEEP SCROLLING
OPEN SOURCE · BSD-2
§ THE STACK YOU BRING
Fork every box.
Pick your inbound channel, STT, LLM, TTS and telephony. Or skip the cascade entirely and switch the whole pipeline to speech-to-speech.
- Swap any provider without rewriting the agent.
- Run hosted, open-weights or air-gapped models.
- Keep the workflow as code you can review.
CLICK ANY BOX TO SWAP
OUTBOUND CALL+ OUTBOUND · WIDGET
S2SGPT Realtime 2 · SPEECH-TO-SPEECH
STTCartesia · SPEECH-TO-TEXT
LLMDograh · BRAIN
TTSDeepgram · TEXT-TO-SPEECH
SIPPlivo · TELEPHONY
PRODUCTION AGENT · READY
§ INTEGRATIONS
BRING YOUR OWN STACK · NO LOCK-INOPENAIGEMINIGOOGLEGROQOPENROUTERAZUREAWSHUGGING FACEMINIMAXSARVAMGLADIAASSEMBLYAISPEECHMATICSDEEPGRAMELEVENLABSCARTESIAINWORLDCAMB.AIRIMESPEACHESSMALLEST AIXAIPLIVOTELNYXTWILIOVOBIZVONAGECLOUDONIXASTERISKLANGFUSETUNEROPENAIGEMINIGOOGLEGROQOPENROUTERAZUREAWSHUGGING FACEMINIMAXSARVAMGLADIAASSEMBLYAISPEECHMATICSDEEPGRAMELEVENLABSCARTESIAINWORLDCAMB.AIRIMESPEACHESSMALLEST AIXAIPLIVOTELNYXTWILIOVOBIZVONAGECLOUDONIXASTERISKLANGFUSETUNER
01 · ON-PREM · DATA RESIDENCY
Your data never leaves.
Deploy Dograh on-prem or inside the VPC of your own cloud. Even the AI models can run within your perimeter or fully air-gapped, so no calls, recordings, transcripts or model inference cross the boundary.
HIPAAGDPRSOC 2ISO 27001CCPA
YOUR DATA BOUNDARYON-PREM · YOUR CLOUD
- voice calls
- recordings
- transcripts
- prompts
- customer PII
- models served in-perimeter
- fully air-gapped
NEVER LEAVES · NEVER TRAINED ON
02 · MODEL CONTEXT PROTOCOL
Build by writing.
Dograh ships an MCP server. Point Claude Code, OpenClaw, Cursor, Codex or any agent runtime at it: they can spin up, modify and deploy complete voice agents without leaving the IDE.
CLAUDE CODE · MCP · DOGRAH● connected
> Build me an AI calling agent for EMI reminders and collections with Dograh.
✓ created agent: emi_collections_v1
✓ generated: reminder + soft-collection scripts
✓ wired: twilio · webhook · pre_call_fetch
> Add a CRM fetch step before the call begins.
✓ added node: pre_call_fetch
✓ wired to: llm.context
_
03 · 2X CONVERSIONS · PRE-RECORDED + TTS
2x conversion. More humane.
Real human voice clips, mixed with TTS in the same cloned voice. The LLM picks a pre-recorded line when one fits and falls back to TTS only when it does not.
HUMAN · PRE-RECORDED0.0s
TTS · LIVE, SAME VOICE0.3s
04 · SPEECH-TO-SPEECH
Real-time. Audio-to-audio.
Audio goes in, audio comes out. No transcription cascade, no text round-trip. Real turn-taking, real interruption handling, ultra-low latency.
Use Gemini 3.1 Flash Live, Vertex AI or GPT Realtime 2.
CASCADE
STT→LLM→TTS
3 HOPS · SLOWERSPEECH-TO-SPEECH
S2S
1 HOP · FASTER- REAL TURN-TAKING
- REAL INTERRUPTION
- 70+ LANGUAGES
05 · SELF-HOSTED STACK · DATA SOVEREIGNTY
Run the whole stack on your infra.
Full data sovereignty, with lower latency as a bonus. Plug in Voxtral, Kokoro, Whisper or whatever runs in your VPC.
WhisperOPENAI · OSS
VoxtralMISTRAL · OSS
Canary QwenNVIDIA · OSS
KokoroOSS
ChatterboxOSS
your voice cloneYOUR WEIGHTS
… AND ANY MODEL YOU CAN HOST
06 · THREE WAYS TO RUN
Same stack. Where you want it.
Let us run it.
Our fully managed cloud, same stack, zero ops. We host, patch and scale. You build agents and ship them.
Your VPC, our ops.
We deploy the entire stack inside your cloud. Data stays on-premise while we handle orchestration, upgrades and reliability.
Talk to founders07 · INDUSTRIES · JURISDICTIONS
Built for the regulated world.
Slots inside your existing compliance perimeter. No HIPAA review on us, no GDPR sign-off on us, no SOC 2 on us. The code is open, so your security team can audit every line before a call goes out.
FintechHealthtechTelemedicineInsuranceBankingLegal intakePharmaDefenseGovernmentEU / GDPRSwitzerlandUKSaudi ArabiaUAESingaporeSouth KoreaJapanUSACanada
08 · QUESTIONS WORTH ASKING
Things worth asking.
Open-source voice infrastructure. Fork the stack, plug in your own models, deploy anywhere.
Why open source?
Three reasons. Your data never leaves your premises, which matters for fintech, healthtech, defense and EU/GDPR teams. You can read, audit and fork every line with no vendor lock-in. And you do not need us to become your compliance vendor, because we never see your calls.
How does the MCP integration work?
Dograh ships a Model Context Protocol server. Point Claude Code, OpenClaw, Cursor, Codex or any MCP client at it and they can create, modify and deploy a complete voice agent without leaving the IDE.
What's the latency like?
The speech-to-speech path is one hop: audio in, audio out. Cascaded STT → LLM → TTS is also supported when you want independent control over each provider.
Can I use my own STT and TTS?
Yes. The workflow builder lets you fork every box. Use a hosted provider, an open-weights model in your VPC, or a mix of both.
Do you support multiple languages?
Yes. Speech-to-speech supports 70+ languages, and self-hosted STT/TTS models can be selected per deployment or per agent.
Can I integrate with my CRM, Slack, Calendly, or webhooks?
Yes. Pre-call fetches, webhooks and post-call actions connect the voice agent to your existing systems without a middleware layer.
What about compliance: HIPAA, GDPR, SOC 2?
Self-hosting keeps calls, recordings, transcripts and prompts inside your existing compliance boundary. Your certifications and controls continue to cover the data because there is no third-party vendor in the path.
How is this different from Vapi, Retell, or LiveKit?
Dograh is a self-hostable full stack with an open workflow builder. You get the orchestration layer, model routing, telephony and operations tooling, and you can run all of it on your own infrastructure.
VOICE INFRASTRUCTURE · OPEN SOURCE · BSD-2
Run voice agents on your own terms.
Start a call in the managed cloud, or fork the repo and run the whole stack yourself. Your data, your models, your perimeter.
Fork every box 70+ languages Self-hosted or managed