If you want a voice agent platform you deploy on your own servers without being locked in, the first job is defining what 'completely on-premises' means — because custody, local inference, and air-gapped are three different things. This guide covers the seven ways voice platforms lock you in, four real options from open-source builds to licensed self-hosted platforms, what a fully local stack costs you in operations, and the exit test that tells you whether you are actually free.
Utkarsh Mohan
Published: Oct 12, 2026
Teams searching for a voice agent platform they can deploy on their own servers are usually carrying two worries at once. One is data: customer conversations contain account details, health information, or commercial secrets, and the team does not want them sitting in someone else's multi-tenant cloud. The other is dependency: they have watched a per-minute price rise, a feature disappear, or a vendor change its terms, and they do not want their phone channel held hostage to a roadmap they do not control.
Those are good instincts, and the market does offer answers. But the phrase "self-hosted" is used loosely enough that two products with the same label can differ by an order of magnitude in how much data actually stays inside your walls. This guide defines the terms, shows where lock-in really hides, lays out the four routes that exist, and is candid about the cost of each — including where Ringlyn's own self-hosted licence sits and where it does not. If your concern is more about regional residency for a regulated finance team, the private cloud and data residency guide is the better starting point.
A voice call is not a database row. It is live audio crossing a carrier network, passing through speech recognition, a language model, and speech synthesis, and then being stored. "Completely on-premises" can mean any of three different things, and each demands a different architecture.
| Tier | What stays inside your boundary | What still leaves | Typical use |
|---|---|---|---|
| 1. Data custody | The application, database, recordings, transcripts, and customer records | Audio and text sent to third-party speech and language providers; telephony carrier traffic | Firms that need records and history under their own control, keys, and retention policy |
| 2. Local inference | Everything in tier 1, plus speech recognition, the language model, and synthesis running on your own GPUs | Only the call itself, between the carrier and your edge (SIP trunk or session border controller) | Organisations that cannot send customer audio or text to any third-party AI provider |
| 3. Air-gapped | Everything in tier 2, plus telephony termination on your own network with no outbound internet dependency | Nothing, apart from your own carrier handoff | Defence, sovereign, and classified environments; contractual no-egress mandates |
Three meanings of on-premises voice AI
Two honest observations follow. First, because a phone call originates on a carrier's network, the audio always crosses a third party before it reaches you; "completely" realistically means from your network edge inward. Second, most commercial products marketed as self-hosted deliver tier 1 and let you reach tier 2 only by connecting models you host yourself. Decide which tier is your actual requirement before you start comparing vendors, because tier 1 and tier 3 are different projects.
Self-hosting removes some lock-in and leaves the rest. Hosting a closed product on your own server still binds you to its runtime, its configuration format, and its licence. Before you commit, test each of these seven points, because a platform can be self-hosted and still be hard to leave.
| Lock-in point | How it traps you | How to test it |
|---|---|---|
| 1. Orchestration runtime | Your call logic only executes inside the vendor's engine, so replacing the vendor means rebuilding every agent | Ask whether the agent definition can be exported and what other system could execute it |
| 2. Prompts and flow definitions | Prompts, tools, and conversation flows live in a proprietary format or a visual builder with no export | Request a full export of one real agent and read it |
| 3. Phone numbers | Numbers are held in the vendor's carrier account, so porting them is a project | Confirm numbers sit in an account you own, with your carrier |
| 4. Voices | Cloned or custom voices exist only on the vendor's synthesis service | Ask whether voice assets can be exported or recreated with another provider |
| 5. Model and provider bundling | Speech and language usage is resold at a markup and cannot be pointed at your own accounts | Check whether you can use your own provider keys and see raw usage |
| 6. Data and history | Call history, transcripts, and analytics can only be read through the vendor's interface | Run a bulk export of transcripts and metadata before you sign |
| 7. Licence and pricing terms | Per-minute pricing, usage telemetry, remote licence checks, or a shut-off if you stop paying | Read what happens to the running system if you cancel, and what connects out |
Seven lock-in points and how to test each one
Notice that point 7 is where "self-hosted" products most often disappoint. A deployment that must phone home for licence verification or metering is on your hardware but not entirely under your control, and a deployment that stops working when a subscription lapses is a lease with extra steps. Ask directly.
Frameworks such as Pipecat and LiveKit Agents give you the real-time plumbing — audio transport, turn-taking, interruption handling, and a pipeline for speech, language, and synthesis — as open-source code you run on your own servers. Pipecat is documented as BSD-2 licensed and free to use with any provider and hosting environment, and LiveKit documents self-hosting its server and agents on your own infrastructure under an open-source licence. Lock-in is minimal because the code is yours and every provider is swappable. You then choose speech and language providers, which means tier 1: your orchestration and storage are yours, and audio goes to whichever managed models you select.
The cost is engineering. You build and maintain the agent, telephony integration, call logging, a dashboard, evaluation, and the operations around them. This is a legitimate choice for a team with voice engineers; the step-by-step build tutorial and the tech stack guide walk through it.
The same frameworks, but with speech recognition, the language model, and synthesis all running on GPUs you own. Common components are open speech-recognition models such as Whisper or faster-whisper, a serving layer such as vLLM, and an open synthesis model. This is the route to tier 2 and, with your own SIP edge and no outbound dependencies, tier 3. It gives you the strongest data position and the least vendor dependency of any option.
It also gives you the most to own. Each model carries its own licence — vLLM being Apache 2.0 says nothing about the licence of the weights you serve with it, and voice packs for synthesis can carry separate terms — so read every licence file rather than the project's headline. And you take on the quality gap, latency tuning, capacity planning, and patching described below.
A finished product — console, agent builder, call logs, analytics, integrations — that you license and deploy into your own server or cloud account. This is the fastest route to production and the lowest staffing requirement, and the best fit when your need is custody of recordings and customer records plus cost predictability. Ringlyn's self-hosted licence is in this category. Lock-in depends on the licence terms: whether you can use your own provider accounts, whether the instance keeps running without a subscription, and whether there is source access. Verify each in writing.
Some SaaS platforms offer an on-prem or customer-cloud deployment to large enterprises. Vapi's documentation, for example, describes deploying its voice infrastructure in your own cloud on your own GPUs, in any location, and notes the deployment regularly sends performance and usage information to Vapi's cloud for GPU adjustment and billing. Rasa states that it can run on-prem, in a private cloud, or air-gapped. Cognigy's current documentation says on-premises installations are no longer offered to new customers. These programmes are usually negotiated, priced for enterprises, and shaped by what the vendor is willing to run outside its own cloud — so ask what exactly runs locally, what connects out, and what the commercial terms are if you leave.
| Dimension | Open-source + managed models | Fully local open-source | Licensed self-hosted platform | Vendor on-prem programme |
|---|---|---|---|---|
| Highest tier reachable | Tier 1 | Tier 3 | Tier 1; tier 2 only if local models are supported | Tier 1 to 2, as negotiated |
| Lock-in risk | Very low | Very low | Low to moderate, depends on terms | Moderate to high |
| Time to production | Weeks to months | Months | Days | Weeks to months |
| Team required | Voice and platform engineers | Engineers plus GPU operations | Light; vendor assists with deployment | Platform team plus vendor management |
| Product surface included | You build it | You build it | Console, analytics, integrations | Full vendor product |
| Cost shape | Engineering time plus usage | Engineering plus GPU capex | Licence plus usage at cost | Enterprise contract plus usage |
Ringlyn's self-hosted licence deploys the console, database, recordings, and transcripts on your infrastructure, with your own telephony and model accounts.
If your requirement is tier 2 or tier 3, you are assembling a real-time system, not installing an application. These are the layers and the decisions inside each.
| Layer | What it does | What you decide |
|---|---|---|
| Telephony edge | Terminates SIP from your carrier and handles media | Which SIP trunk or session border controller; where it sits; how failover works |
| Real-time orchestration | Moves audio, detects speech and turn-taking, handles interruptions | Framework (Pipecat, LiveKit Agents, or your own); how you handle barge-in and silence |
| Speech recognition | Turns caller audio into text in real time | Model and size, streaming behaviour, language coverage, accuracy on your callers' accents and audio quality |
| Language model | Decides what to say and which tools to call | Model, size, quantisation, context length, tool-calling reliability, and its licence |
| Speech synthesis | Turns responses into natural audio | Voice quality, latency to first audio, licence on the voice |
| Storage and retention | Recordings, transcripts, structured outputs | Encryption, retention schedule, legal hold, deletion |
| Observability and evaluation | Tracing, call QA, regression tests | How you test a prompt change before it reaches production calls |
| GPU capacity | Runs the three models concurrently under peak load | Hardware, concurrency headroom, redundancy, and a plan for a node failure |
The layers of a fully local voice agent stack
Three costs are routinely underestimated. Quality: the best hosted models are generally ahead of what you can run locally on a given budget, particularly for tool-calling reliability and handling unscripted conversation; test your own call scenarios rather than trusting a leaderboard. Latency: callers notice delay quickly, and every stage adds to it, so the whole chain must be tuned together rather than model by model. Capacity: concurrency per GPU depends on model size, prompt length, and your target latency, so do not rely on anyone's published per-GPU call count — load test with your own prompts and peak traffic, and budget redundancy.
A sensible approach for many teams is to run locally only what the requirement forces. If speech recognition and the language model see the sensitive content, host those; synthesis, which mostly speaks your scripted wording, is the lowest-risk component to leave managed if your policy allows. That hybrid keeps your data position strong while limiting what you operate. The enterprise compliance guide covers the component-by-component hybrid pattern in more depth.
Ringlyn's self-hosted licence is a tier 1 product in this framework, and it is worth being exact about that. Per the self-hosted documentation, the application and admin console, the database and object storage, and all call records, transcripts, and recordings run inside your own server or cloud account, and nothing is mirrored to a Ringlyn-hosted production system afterwards. Telephony and the speech and language models are connected through your own provider accounts and API keys, and the documented infrastructure requirements name Twilio for telephony, Gemini for AI, and Supabase for the database.
| Lock-in point | Ringlyn self-hosted licence position |
|---|---|
| Orchestration runtime | The platform runs the agents; there is no source code under the self-hosted licence, so the runtime is Ringlyn's |
| Phone numbers | Held in your own Twilio or SIP provider account, so you control routing and porting |
| Model and provider bundling | Your own provider accounts and keys; you see the usage and pay the providers directly, with no markup in between |
| Data and history | In your database and storage, so exports are queries you run |
| Licence and pricing | One-time $2,500 licence, then $500 a year for updates and support from year two; if you cancel, the deployed instance keeps running and is not remotely disabled |
| Source access | Not included. The Agency Licence includes the full production source code and resale rights |
Ringlyn self-hosted measured against the lock-in points
So it removes per-minute platform fees, keeps your records and keys under your control, and leaves you free to switch telephony and model providers. It does not, on its own, remove third-party processing: audio and text sent to the providers you connect leave your network. If your requirement is that nothing reaches a third-party AI service, ask Ringlyn directly whether pointing the platform at locally hosted speech and language models is supported for your deployment, and what the telemetry and licence connections are, before relying on it. And if source access is what lock-in means to you, the Agency licence is the one that includes it; the agency versus self-hosted comparison sets out the difference.
| If your requirement is | Start with | Why |
|---|---|---|
| Recordings and customer records must sit in an environment we control, with our own keys and retention | Licensed self-hosted platform | Fastest route to custody with the least staff |
| We have voice engineers and want to own the code and swap any provider | Open-source framework with managed models | Maximum flexibility with an engineering team you already have |
| No customer audio or text may reach a third-party AI provider | Fully local open-source stack, or a platform confirmed to support local models | Only local inference removes third-party processing |
| Air-gapped or sovereign environment with no outbound internet | Fully local open-source stack with your own SIP edge | Few commercial products ship fully offline; verify every connection |
| We already have an enterprise agreement with a SaaS vendor | That vendor's on-prem programme, with its egress list reviewed | Lower switching effort, but check telemetry and exit terms |
| We want to resell voice AI under our own brand | Agency source-code licence | Includes source, rebranding, and resale rights |
Choosing a route by requirement
The most reliable way to measure lock-in is to ask what leaving would take, and to prove the answers before you sign rather than after. Run this test during evaluation:
Decide first which on-premises you mean: custody of your data, local inference, or fully air-gapped. A licensed self-hosted platform gets you custody quickly and cheaply; an open-source build with local models gets you the strongest data position and the least lock-in at the price of an engineering team and GPU operations; vendor on-prem programmes sit between, with connectivity footnotes worth reading. Whichever route you take, test lock-in at the seven points, get the egress list, and prove your exit before you sign. If custody and cost predictability are what you need, Ringlyn's self-hosted licence is built for exactly that, with the third-party boundary described honestly above.
A self-hosted voice agent platform is one you deploy and operate in your own server or cloud account rather than consuming as a multi-tenant SaaS. The application, database, call recordings, and transcripts live in your environment under your security controls. Telephony and speech or language models are typically connected with your own provider accounts, unless you host those models yourself as well.
Test seven points before you sign: whether agent definitions can be exported and executed elsewhere, whether prompts and flows are readable, whether phone numbers sit in your own carrier account, whether voices can be recreated, whether you can use your own provider keys, whether call data can be bulk exported, and what happens to the running system if you cancel. Open-source frameworks such as Pipecat and LiveKit Agents minimise lock-in because the code is yours.
Yes, but only if speech recognition, the language model, and speech synthesis also run on your own infrastructure, and your telephony terminates on your own network edge. Many self-hosted platforms keep the application and records on-premises but still send audio and text to third-party model providers. Confirm exactly which components run locally and which connect out.
Pipecat and LiveKit Agents are open-source frameworks for the real-time pipeline, and both can be self-hosted. For fully local inference, teams commonly pair them with open speech-recognition models such as Whisper or faster-whisper, a serving layer such as vLLM, and an open speech-synthesis model. Check the licence of every model and voice you use, because each carries its own terms.
It keeps the application, database, recordings, transcripts, and customer records in your environment, with nothing mirrored to Ringlyn's production systems. Telephony and the speech and language models are connected with your own provider accounts, and the documented requirements name Twilio, Gemini, and Supabase, so audio and text sent to those providers are governed by their regions and terms. Ask Ringlyn directly about local-model support if you need zero third-party processing.
The deployed instance keeps running; you simply stop receiving updates and support, and the platform is not remotely disabled. The licence is $2,500 one-time with $500 a year for updates and support from the second year. It does not include source-code access or resale rights, which are part of the Agency Licence.
Often at higher call volumes, because the platform fee becomes fixed and only telephony and model usage scale with calls. Ringlyn's own modelling puts break-even for a licensed self-hosted deployment somewhere between month six and month twelve at around 2,000 minutes a month and within the first quarter at around 10,000. A fully local open-source build adds engineering and GPU costs that can outweigh the savings at low volume.
It depends on the models you choose, prompt length, your latency target, and peak concurrent calls, so no published per-GPU figure should be trusted. Load test with your own prompts and realistic peak traffic, keep headroom, and plan for the failure of a node. Running speech recognition, the language model, and synthesis concurrently is what drives the sizing.
A fully local open-source stack with your own SIP edge is the most reliable route, because few commercial products ship fully offline. If you consider a commercial platform, request the full list of outbound connections, confirm licence verification and telemetry can be blocked, and test the system in an environment with no internet access before you commit.
The seven data flows a residency guarantee must cover, four deployment models, and ten questions that expose a vendor's real position.
A decision framework covering ownership, cost curve, compliance posture, resale rights, and operational load.
Build a production-ready AI voice agent: STT, LLM, TTS, telephony, and real-time orchestration with Pipecat.