// the find
dograh-hq/dograh
Open source voice AI platform. Self-hosted alternative to Vapi and Retell. On Prem, BYOK across Speech to Speech or LLM/STT/TTS, with a visual workflow builder, MCP native and telephony support.
Dograh is a self-hosted voice agent platform — think Vapi or Retell but open source and running on your own infra. It wraps Pipecat with a visual workflow builder, handles the telephony plumbing (Twilio, Telnyx, Asterisk ARI, and more), and lets you swap in whatever LLM/STT/TTS stack you want. The target is teams that need voice AI in production but can't hand call data to a SaaS vendor.
The BYOK and BYOL story is genuinely complete — you can mix Deepgram STT with ElevenLabs TTS and a local LLM, and the configuration layer handles merging and masking secrets sanely. Telephony coverage is broad: seven providers including Asterisk ARI for on-prem PBX integration, which no SaaS alternative touches. The MCP server is a real differentiator — letting a coding agent inspect and edit workflows programmatically is a concrete workflow, not a feature flag. Migration history is extensive (80+ Alembic migrations) and the schema is clearly production-hardened, not a prototype.
The architecture leans heavily on Pipecat, which means you're inheriting Pipecat's latency characteristics and frame pipeline model — if Pipecat has a bad release, so does Dograh. The campaign orchestrator has a circuit breaker and rate limiter but the README says almost nothing about how outbound campaigns actually scale under load, which is the hard part. There's a Node/TypeScript validator tucked inside the Python MCP server (`ts_validator/`) that gets shelled out via a bridge — that's an awkward seam that will cause pain when Node versions drift. Self-hosting is Docker Compose only; if you need Kubernetes or anything beyond a single-node setup, you're writing that yourself.