Dograh

What Dograh is

Dograh is an open-source voice agent platform. Its no-code page puts it plainly: "Dograh is a visual workflow builder for voice agents. Drag, connect, and ship." The homepage calls the whole thing "Open-source voice infrastructure", and a longer article on its site describes it as "a visual workflow builder (think n8n, but for voice agents)" released under a BSD 2-Clause license on GitHub. It speaks to three readers at once: product and operations people who want to edit a call flow without a developer, engineers who want to fork and run the stack themselves, and security or compliance teams that cannot send call data to a hosted vendor. The pages do not say where the company is based.

What you build and how

The editor is a drag-and-drop canvas. "Design every conversational state, branch, and tool call visually, then ship." The page argues that code-first voice stacks become a bottleneck, since "Every prompt tweak needs a PR. Every new branch needs an engineer." In Dograh, "Product, ops, and engineering all edit the same flow, the diff is the diagram". It lists multi-agent flows too: "Each agent owns its slice of the conversation, with clear handoffs between them."

Testing happens in the same tool. You "Test conversations live inside the builder, listen to recordings, and refine prompts in place."

Speech and models. The page tags the audio side as the dependable part and the model side as the experimental part: "STT and TTS are production-ready. LLM swapping is in beta." The speech-to-text options shown are Whisper, Voxtral and Canary Qwen, all labeled open source. The text-to-speech options are Kokoro and Chatterbox, "your voice clone", and "any model you can host". The no-code page adds that you can plug in providers such as Deepgram and ElevenLabs. Bringing your own language model, including open-source ones, is described as "experimental today, fine for power users testing reasoning backends".

Voice techniques. The homepage describes a hybrid in which real pre-recorded human clips are mixed with synthetic speech in the same cloned voice. "The LLM picks a pre-recorded line when one fits, and falls back to TTS only when it doesn't." It claims "2x conversion" for this and cost "down up to 3×". It also describes speech-to-speech mode: "Audio goes in, audio comes out. No transcription cascade, no text round-trip."

Building by writing. Dograh "ships a Model Context Protocol server". You point Claude Code, OpenClaw, Cursor or "any agent runtime" at it, and they can "spin up, modify, and deploy full voice agents on your stack, without leaving the IDE". The page's sample session asks for an agent that handles EMI reminders and collections, and the tool creates the agent, generates reminder and soft-collection scripts, and wires Twilio, a webhook and a pre-call CRM fetch step. A later request adds a CRM fetch node before the call begins. This is a second way to build the same workflow, alongside the canvas.

Integrations. "CRMs, telephony, analytics, and arbitrary APIs all connect through the builder." The only specific names on the pages are Twilio and webhooks in that sample, plus Deepgram and ElevenLabs on the speech side. The homepage FAQ lists questions about CRM, Slack, Calendly and webhooks, and the answers are not shown, so check them for any integration you depend on. The site also carries many solution pages, with titles such as appointment booking, lead generation, a virtual receptionist for small business, healthcare and pharma, debt collection, and a HubSpot CRM integration.

Channels. Everything described is voice. There is no text chat bot, which is why the chatbot builder listing here rests on the visual conversation builder rather than a messaging channel.

Running it. "Three ways to run": self-host the open-source stack ("Clone the repo, docker compose up, you're running. Forever free, BSD 2-Clause, your servers, your data"), use the managed cloud ("We host, patch, and scale"), or have the company "deploy the entire stack inside your VPC" and operate it. The self-hosting prerequisite is Docker 20.10 or newer, and the article says it deploys "in under 2 minutes". For open-source low-code, this is the core appeal: you can read, fork and modify the code, and nothing locks you to the hosted service.

The evidence from the pages

• License. BSD 2-Clause, stated on the homepage and in the article. The code lives on GitHub, and every page carries a "Fork on GitHub" button.
• Price. Self-hosting is "Forever free". The cloud has a "Try Cloud free" button. No paid plan name, per-minute rate or usage limit appears on the pages, and the managed private cloud is a "Talk to founders" route. The article compares hosted rivals on their prices, and those figures belong to those products.
• Data and compliance. The homepage says "no calls, recordings, transcripts, or model inference ever leave your boundary", with a fully air-gapped option. Read the compliance wording closely. Dograh does not claim HIPAA, GDPR, SOC 2 or ISO certification for itself. It says the data stays in your perimeter "so your existing certifications and controls keep covering it", and that "you don't need us to be SOC 2 or HIPAA or GDPR-certified because we don't see your data." That holds for the self-hosted and private-cloud options. For the managed cloud, the pages do not say what covers your data.
• Sectors and regions. The page lists fintech, healthtech, telemedicine, insurance, banking, legal intake, pharma, defense and government as sectors, and the EU, Switzerland, the UK, Saudi Arabia, the UAE, Singapore, South Korea, Japan, the USA and Canada as jurisdictions. It is a marketing list of where the open-source model could fit, not a customer list.
• Support. Documentation, Slack discussions, a community and a founders email address are linked.

Who it suits and who it does not

Dograh suits a regulated team that must keep voice data inside its own infrastructure, a developer who wants a no-code builder plus an escape hatch into the source, and a team of mixed skills where operations edits the flow and engineers own the deployment. The MCP server will appeal to people who already work in an AI coding tool.

Caveats are real. Self-hosting means you run it: you patch it, scale it and choose your speech models. The company itself calls LLM swapping beta. The pages describe no chat, email or SMS channel, and no published pricing for the cloud. A small business that only wants an answering line live this afternoon will find a hosted receptionist product simpler.