Four vendor invoices arrive the month you run an AI voice agent, regardless of which path you take to get there. Telephony: Twilio outbound local at $0.0140/min plus Media Streams at $0.0044/min (twilio.com/en-us/voice/pricing/us, 2026-08-07). STT: Deepgram Nova-3 at $0.0048/min on pay-as-you-go (deepgram.com/pricing, 2026-08-07). TTS: Cartesia-class voice at approximately $0.015/min; ElevenLabs-class voice at approximately $0.040/min (market rates from retellai.com/pricing, 2026-08-07). LLM: GPT-5-nano-class at $0.003/min through GPT-4.1-class at $0.045/min, from model provider rates via retellai.com/pricing, same date.
On a mid-range outbound stack — GPT-4.1-class model, Cartesia-class voice, standard Twilio telephony — those four layers total roughly $0.068 per minute. At 10,000 minutes a month, that is $680, before any platform, before any admin panel, before any multi-tenant billing layer. No architecture removes this floor. What every platform in this category — hosted or self-hosted, rented or purchased — charges on top of it is the variable this page compares. Callia is a multi-tenant AI voice agent platform sold once as source code. You deploy it under your own brand, connect your own provider keys across five layers (telephony, STT, TTS, LLM, and middleware), and bill your own clients' usage through your own Stripe account. This page compares it against Dograh — the self-hosted alternative in the same "deploy on your own infrastructure" position — and the four routes to becoming an AI voice-agent operator.
Decision guide
| If the requirement is… | The honest answer |
|---|---|
| Running voice agents for your own organisation's calls — zero platform licence cost | Dograh — BSD-2-Clause, $0, dograh-hq/dograh, active 2026-08-08 |
| Building a custom voice pipeline with framework-level control and no admin overhead | Pipecat (BSD-2, 13,978 stars, pipecat-ai/pipecat, active 2026-08-07) or LiveKit Agents (Apache-2.0, 12,720 stars, livekit/agents, active 2026-08-07) |
| Launching immediately on managed infrastructure; the calls are your own | Vapi ($0.05/min + at-cost providers, vapi.ai/pricing, 2026-08-07) or Retell ($0.07–$0.31/min all-in, retellai.com/pricing, same date) |
| White-label managed service with per-client pricing; recurring platform subscription acceptable | Callin.io — approximately $119–$499/mo per comparison sources (not verified against Callin.io's own pricing page; confirm directly before deciding) |
| Own the codebase, deploy under your own brand, bill your own clients, remove the per-minute platform margin | Callia — See the demo → · Buy and deploy → |
Capability comparison
| Callia | Dograh | |
|---|---|---|
| Licence | Commercial (one-time purchase) | BSD-2-Clause ($0, permissive) |
| Provider stack | 5 swappable layers: telephony / STT / TTS / LLM / middleware | BYOK across STT/TTS/LLM or speech-to-speech |
| Visual workflow builder | Yes (Call Flow Designer, drag-and-drop nodes) | Yes |
| MCP server | Yes | Yes |
| Multi-tenant auth + RBAC | Yes (super_admin / admin / member) | Not in scope — single-operator tool |
| DB-driven plans + client billing | Yes (Stripe abstraction, real-time credit metering) | No |
| Outbound campaign dial engine | Yes (QStash-backed, DNC check before every call) | Not in scope |
| Knowledge base with RAG | Yes (pgvector, per-agent ingestion, gap analytics) | Not in scope |
| Call logs + recordings | Yes (transcript, audio player, disposition tagging) | No |
| Analytics | Yes (cost per call, agent comparison, KB gap tracking) | No |
| Commercial operating licence (deploy, brand, charge your own clients) | Yes | BSD-2 permits redistribution; the commercial SaaS layer is absent |
| What Dograh does that Callia does not | — | $0 licence cost, always; BSD-2-Clause permissiveness; active community with near-daily contributions; broader ecosystem reach from community adoption |
Two comparisons follow. The first sets Callia against Dograh, which occupies the same ownership position and the same buyer level: software you deploy yourself, with your own provider keys. The second weighs the four routes to operating a voice-agent platform as a business: rent a hosted service, commission a custom build, start from generic boilerplate, or buy this codebase. Both stay at the operator level — not someone who wants to make a phone call, but someone who wants to own the system that makes them possible.
Callia vs Dograh
What each is, at the owner level
Dograh is a self-hosted AI voice agent platform published on GitHub under the BSD-2-Clause licence. It ships a visual workflow builder, telephony support, BYOK across the full STT/TTS/LLM stack — or speech-to-speech — and an MCP server. Its own repo describes it as "self-hosted alternative to Vapi and Retell." As of 2026-08-08: 5,163 GitHub stars, pushed that day, actively maintained. Licence cost: zero. (dograh-hq/dograh, GitHub API, 2026-08-08.)
Callia is a commercial SaaS platform sold under a one-time commercial licence. The buyer deploys it, brands it, and sells subscriptions to their own clients. The distinguishing layer is the commercial SaaS foundation: auth, RBAC, DB-driven plans, billing abstraction where client subscriptions route to the operator's Stripe account, a credit system metered in call-minutes and deducted in real time, a campaign dial engine with DNC checking before every call, call logs with transcripts and disposition tagging, and an admin panel already wired. That is the platform a voice-AI operator would otherwise have to build themselves before any voice agent runs.
When Dograh is the right call
Dograh is $0, under one of the most permissive licences in existence. If you want to run AI voice agents for your own organisation's calls, extend an active community project, and do not need multi-tenant billing, a resale-ready admin panel, or a commercial SaaS licence to operate a business on, Dograh is a valid and honest answer — and this page is prepared to say so. Its development pace, community contributions, and near-daily activity are real advantages a v1 commercial product cannot yet match. It ships a visual workflow builder and an MCP server, so the overlap with Callia's feature surface is genuine, not marginal.
When owning Callia is the right call
Dograh is built for your own operations. It does not ship the layer that lets you become the platform for someone else's operations — the multi-tenant auth split, the plan engine where you configure pricing for your tenants, the billing abstraction where their subscriptions flow into your Stripe account, a call-minute credit system that deducts in real time per call, a campaign dial engine across a tenant's contact list with DNC enforcement, or a commercial licence purpose-built for operating the platform as a product you sell. Callia is what you deploy when the calls belong to your clients, not to you.
The middleware counterweight — the most important cost honesty on this page
Callia ships two pipeline modes. Middleware Mode is the default and the recommended starting point: calls route through Vapi or LiveKit Agents, the entire application runs on Vercel serverless without any extra service to run, and barge-in is handled automatically. In Middleware Mode, the Callia operator is still paying Vapi's $0.05/min platform fee (vapi.ai/pricing, 2026-08-07). The lower per-minute economics belong to Direct Mode, which routes calls through your own provider stack. Direct Mode requires a separate always-on relay service alongside the Vercel app — because a live-call WebSocket cannot survive serverless function timeouts — estimated at $5–15/month for a solo worker (internal project estimate, not a vendor quote). Both modes ship in the product. The cost difference is real. The infrastructure requirement belongs in the same sentence as the saving, every time.
What self-hosting actually removes — a band, not a single number
Against Retell mid-configuration ($0.130/min, retellai.com/pricing, 2026-08-07): self-hosting in Direct Mode at $0.068/min removes about 48% of the per-minute cost. Against Retell premium with a frontier model and ElevenLabs voice ($0.190/min, same source): self-hosting at ~$0.144/min removes about 24%. Against Bland Start ($0.140/min, bland.ai/pricing, 2026-08-07): lean stack removes about 73%, mid stack about 51%. Against Vapi with your own API keys — where you pay ~$0.068/min in providers plus Vapi's $0.050/min fee, totalling ~$0.118/min — the saving is exactly $0.050/min, the platform fee and nothing else (vapi.ai/pricing, 2026-08-07). Anyone quoting one savings percentage is describing one configuration and one provider stack. The band is 24%–73%. At 10,000 minutes a month, that band translates to saving roughly $460–$1,020/month while still paying roughly $380–$940/month in irreducible provider costs.
The four acquisition paths
These four paths all arrive at the same destination — operating an AI voice-agent platform as a business — and they share the same buyer: someone who wants to own the outcome. The comparison is across acquisition routes, not tools for different jobs.
| Path | Upfront | Ongoing | Infrastructure | Domain layer |
|---|---|---|---|---|
| Rent a hosted platform (Vapi, Retell, Bland, ElevenLabs Agents) | $0 | $0.05–$0.19/min platform margin + provider cost; Synthflow enterprise at $30,000/yr annual minimum as of 2026-08-07 | Platform's | Platform owns it |
| Commission a custom build | Senior engineering time (months) | Engineering maintenance + hosting | Yours | You build it |
| Start from generic boilerplate | Near $0 | Engineering maintenance | Yours | Not included |
| Buy Callia | $349 | Hosting + provider costs; relay worker ~$5–15/mo in Direct Mode | Yours | Ships with it |
Building this category from scratch means implementing a persistent WebSocket relay service (a Vercel-incompatible always-on connection), call flow state management with interruption handling, a per-second credit metering system rounded up to the minute, a multi-tenant plan engine where client subscriptions route to the operator's Stripe account, a knowledge base with per-agent document ingestion and pgvector cosine retrieval, and an outbound campaign dial engine with DNC enforcement before every call. Each of those is its own implementation project and an ongoing maintenance surface. Generic boilerplate provides the multi-tenant shell; the voice-agent domain layer is not included and must be built from scratch.
On the hosted platforms: Vapi already supports BYO keys across transcription, LLM and all voice providers — and charges $0.05/min for operating on their infrastructure regardless (vapi.ai/pricing, 2026-08-07). Retell supports a custom LLM over WebSocket and BYO telephony via SIP trunking — and its voice infrastructure fee is $0.055/min before any model or voice is added (retellai.com/pricing, 2026-08-07). Bland supports BYO Twilio — and charges $0.14/min all-in on its Start tier (bland.ai/pricing, 2026-08-07). ElevenLabs Agents bundles minutes into plan tiers; LLM and telephony are billed separately on top, and even a custom model routes its fees through ElevenLabs credits (elevenlabs.io/pricing/agents, 2026-08-07). None of the four are sold as a platform you white-label, deploy under your own brand, and resell to your own clients. BYO keys ends at the keys. The orchestration, the call data, and the platform itself remain theirs.
Common questions
Questions an operator asks before this decision. FAQPage JSON-LD follows.
Is Dograh free for commercial use?
Yes. Dograh is published under BSD-2-Clause, one of the most permissive licences available. It permits commercial use, modification, redistribution, and sublicensing without copyleft obligations — there is no attribution requirement beyond preserving the licence notice. The scope caveat worth stating plainly: Dograh is a voice-agent deployment tool, not a commercial SaaS platform with multi-tenant billing, a campaign dial engine, and a resale-ready admin panel. Those are different products. If you need the first but not the second, Dograh is the right call and this page is honest about that. (dograh-hq/dograh, GitHub API, 2026-08-08.)
Do I still pay Twilio and my AI providers if I self-host a voice agent?
Yes. Self-hosting removes the platform's orchestration margin — Vapi's $0.05/min or Retell's $0.055/min voice infrastructure fee — but not the four provider bills underneath. On a mid-range outbound stack, telephony plus STT plus TTS plus LLM totals roughly $0.068 per minute, derived from Twilio's, Deepgram's and model vendor pricing pages as of 2026-08-07. At 10,000 minutes a month, that is approximately $680, every month, regardless of which platform owns the layer above it. Self-hosting removes the margin on top of the floor. The floor is not removable.
What is Callia's default pipeline mode, and does it remove the Vapi platform fee?
Callia's default is Middleware Mode, which routes calls through Vapi or LiveKit Agents and deploys entirely on Vercel serverless — no additional service required. In Middleware Mode, the operator is still paying Vapi's $0.05/min platform fee. The lower per-minute economics belong to Direct Mode, which routes calls through your own provider stack. Direct Mode requires a separate always-on relay service alongside the Vercel app — a live-call WebSocket cannot survive a serverless function timeout — estimated at $5–15/month for a solo worker. Both modes ship in the product; the choice is a deploy-time decision the operator makes, not a feature gap.
Can I bring my own Twilio, ElevenLabs, and LLM keys with Vapi or Retell?
Yes, across most layers. Vapi's own documentation states: "Once your API key is validated, you won't be charged when using that provider through Vapi" — across transcription, LLM and all voice providers (docs.vapi.ai/customization/provider-keys, 2026-08-07). Retell supports a custom LLM over WebSocket and BYO telephony via SIP trunking. Bland supports BYO Twilio. BYO keys is not a Callia-specific claim. What BYO keys does not change: Vapi's $0.05/min orchestration fee and Retell's $0.055/min voice infrastructure fee persist regardless of whose keys are in use. Those fees are the cost of operating on their infrastructure, not a markup on your providers.
What does Synthflow cost in 2026?
As of 2026-08-07, Synthflow's public pricing page shows an enterprise engagement starting at $30,000 annually, with no self-serve tier and a "Contact Sales" path for everyone else. The pay-as-you-go component rates previously documented in Synthflow's own documentation are no longer accessible at a primary URL — that documentation page now redirects to the enterprise page. Any per-minute or per-component figures from secondary sources are not confirmable against Synthflow's own current page. Verify directly with Synthflow before relying on any secondary-source figure.
Are Pipecat or LiveKit Agents usable alternatives for running a voice-agent platform?
Both are frameworks, not finished products. Pipecat (BSD-2-Clause, 13,978 GitHub stars, pushed 2026-08-07, pipecat-ai/pipecat) gives you a Python pipeline of frame processors you assemble yourself. LiveKit Agents (Apache-2.0, 12,720 stars, pushed 2026-08-07, livekit/agents) builds on LiveKit's WebRTC server with turn detection, interruption handling, and telephony from the same source. Neither ships auth, RBAC, multi-tenancy, billing, a plan engine, an admin panel, call logs, or campaign management. They are the right answer when the requirement is building a custom voice pipeline from its components; they are not the right answer when the requirement is buying a finished platform to operate and bill through.