dograh

Building a voice agent that can hold a phone conversation, qualify a lead, or route a caller used to mean either writing low-level audio pipeline code or renting a managed platform like Vapi or Retell. The managed route is fast but ties you to per-minute platform fees on top of your model bills, and your call data and transcripts live on someone else's infrastructure. dograh, from dograh-hq, is an open-source, self-hostable platform for building production voice agents for both inbound and outbound calls, and it positions itself directly as a self-hosted alternative to Vapi and Retell. The core idea is a visual, drag-and-drop workflow builder so that designing an agent's conversation logic does not require a code-first framework, while you keep full control of the deployment, the data, and the model providers underneath.

Under the hood dograh is built on Pipecat, the open-source real-time voice pipeline, which it wraps in a workflow editor and an operations layer. It follows a bring-your-own-keys model: dograh ships no speech or language engine of its own and instead orchestrates whatever LLM, STT, and TTS providers you plug in, whether a hosted API or a local model. It supports both the classic STT to LLM to TTS chain and speech-to-speech setups. Telephony runs through Twilio, Vonage, Telnyx, and Cloudonix for real inbound and outbound calls, with human-agent transfer. Two things stand out: a QA Node for checking prompt quality inside a flow, and a Test Mode plus in-dashboard web calls so you can talk to the bot before wiring up phone numbers. It is also MCP-native, so agents can call external tools.

dograh is aimed at teams that need to run voice agents at production scale but want data control and the freedom to swap providers as costs and quality shift. The clearest fit is regulated or data-sensitive work: lead qualification, appointment intake, insurance or claims screening, customer-service triage, and outbound follow-up campaigns where transcripts cannot leave your own infrastructure. Because it is BYOK and self-hosted, it suits organizations that already pay model vendors directly and resent an extra per-minute platform margin. The visual builder lowers the barrier enough that a product or operations person can shape conversation logic, while engineers handle deployment and the keys. Python and Node.js SDKs and a managed cloud option at app.dograh.com exist for teams that want to start without standing up their own stack first.

The honest caveats are real. dograh requires a Docker setup, and hardware specifications are not documented, so capacity planning for concurrent calls is something you will have to size and load-test yourself. Because it brings no engine of its own, voice quality, latency, and reliability depend entirely on the LLM, STT, and TTS keys you supply, and validating that an agent behaves safely on live calls is on you. The project reports it is currently English-focused, with other languages described as expandable rather than ready. Telemetry is on by default and must be disabled via an environment variable. As a fast-moving project resting on Pipecat as a submodule, you also take on the maintenance of a moving dependency, and the BSD-2-Clause license, while permissively open, means no warranty for production incidents.

Against Vapi and Retell, dograh trades polish and a fully managed runtime for ownership: you give up the convenience of a hosted control plane and gain source access, data residency, and freedom from platform per-minute fees. For teams that can run Docker, already hold their own model keys, and care about where call data lives, it is a credible self-hosted base, especially with the QA Node, Test Mode, and telephony integrations reducing the gap to a turnkey product. It is too early if you expect plug-and-play out of the box without your own keys, need broad multilingual support today, or lack the operational appetite to load-test and validate reliability yourself. Pilot it on a non-critical flow, measure latency and accuracy with your chosen providers, then decide whether to scale.