Contents
- Verdict
- What is Vocode?
- Best for, and not for
- Strengths and limits
- Pricing
- What people actually build with Vocode - and why most of it has moved
- What Vocode actually costs - and why free is the wrong question
- The real finding: the project is dormant
- What breaks in production, and how you find out
- What it connects to that you already run
- Vocode alternatives: what to pick instead, and when
- Migration and lock-in: can you get your work out?
- FAQ
Vocode is an MIT-licensed open-source library for building voice LLM applications, and the most important thing a buyer needs to know in August 2026 is that it is dormant. We checked the GitHub API directly on 6 August 2026: the main vocode-core repository was last pushed to on 15 November 2024, it carries 3,781 stars and just 2 open issues, and its most recent tagged release, v0.1.114a2, dates from August 2024. No repository in the vocodedev organisation has been pushed to since November 2024. That is roughly twenty-one months of no activity in a field that has changed enormously over the same period. The MIT licence means the code you have cannot be taken away, but an unmaintained voice library is not a foundation for a business process, and we would not recommend starting a new build on it today.
What is Vocode?
Vocode is an open source Python library for building voice-based LLM applications, one of the earliest projects to make real-time voice agents accessible to developers. It provides abstractions over speech-to-text, LLM, and text-to-speech providers, plus telephony support via Twilio, so you can compose a voice agent from vendors you choose and self-host the whole thing. It was historically influential in the space, but active development has slowed dramatically since late 2024.
Best for
Developers who want a free, self-hosted starting point to learn the voice agent stack and do not mind maintaining the code themselves.
Not for
Any business without a Python developer, and any process where reliability matters more than control. An open-source library with no maintainer activity since November 2024 is a foundation you inherit, not a supported product.
Strengths
- MIT licensed, which is the most permissive position available - nobody can change the terms on code you already have
- Free to self-host, so your only costs are the providers you would pay anyway
- Full control over every component: transcription, model, voice and telephony are all your choice
- Deploys to phone calls, video meetings and other channels rather than being telephony-only
- 3,781 GitHub stars and a body of existing example code, so the material for learning from it is there
- Because it is a library rather than a platform, it integrates into an existing Python application naturally
Limitations
- Dormant: no push to any repository in the vocodedev organisation since 15 November 2024, and no tagged release since August 2024, re-verified via the GitHub API on 6 August 2026
- Only 2 open issues against 3,781 stars, which on a project this size indicates people have stopped filing rather than that the backlog was cleared
- It is a library, so everything a product would give you - dashboard, monitoring, support - is yours to build or do without
- Provider SDKs and model APIs have moved on considerably since the last release, so expect to fix breakage yourself before anything works
- We could not verify current hosted pricing from a public page, so the managed route cannot be budgeted from published figures
- Latency tuning across four providers is genuinely hard, and it is the difference between a usable agent and an irritating one
- No packaged compliance story - if you need HIPAA handling you design and evidence it yourself
- Requires Python engineering capacity to build and, more importantly, to keep running
Vocode pricing
Vocode's core library is MIT licensed and free - you pay nothing for the software and everything for the providers you assemble behind it: speech-to-text, language model, text-to-speech and telephony, each billed by its own vendor at its own rate. A hosted service exists alongside the open-source core, but we could not verify its current pricing from a publicly readable page as of August 2026, so we are not quoting figures. For a self-hosted deployment, the realistic monthly cost is the sum of your provider bills plus a small server: roughly $60 to $150 at 1,500 minutes a month, or about $136 at the category reference workload of 4,000 minutes, depending entirely on which model and voice you choose.
| Plan | Price | What you get |
|---|---|---|
| Open source (self-host) | $0 | MIT licensed core library; you pay only your own providers and hosting |
| Provider costs | Your own accounts | STT, LLM, TTS and telephony each billed directly by their vendor at their own rates |
| Hosted offering | Not verified | A hosted service exists; we could not verify current pricing from a public page as of Aug 2026 |
| Worked example (1,500 min) | ~$60-$150 | 1,500 minutes a month self-hosted: provider costs plus a small server, driven mainly by model and voice choice |
| Worked example (4,000 min) | ~$136 | At the category reference workload: Deepgram STT and TTS, GPT-5 mini and Twilio, plus a small server. Excludes the engineering time to revive the library |
Pricing reflects public plans as of August 6, 2026 and can change. Check Vocode for the latest.
What people actually build with Vocode - and why most of it has moved
Vocode was a genuinely good idea executed early: an MIT-licensed Python library that let you assemble a voice agent from your own transcription, model, speech and telephony providers and drop it inside an application you already had. The use cases below are the ones it was built for and the ones you will find in its examples directory. The honest framing in 2026 is that they are all still sensible things to build and Vocode is no longer the sensible way to build them, because the library has not been touched since November 2024.
What the library was used for
Embedding voice inside an existing Python service. A team already running a Django or FastAPI application could add a phone agent as an import rather than as an external service, sharing the same database session, the same models and the same authentication. This is the one case where a library genuinely beat a platform, and it remains the only real argument for Vocode's architecture.
Outbound survey and research calls. A script iterates a list, places calls, asks a fixed set of questions, and writes structured answers back to a database. Volume is bursty, the conversation is shallow, and per-minute platform fees on thousands of short calls added up in a way that a self-hosted library did not.
Internal voice tooling. An on-call escalation line that reads out the current incident status, or a warehouse line that looks up stock by spoken SKU. Nobody external hears it, reliability requirements are modest, and the team is happy to fix it themselves. This is the category where an unmaintained dependency is genuinely acceptable.
Prototyping a voice product before committing to a vendor. Because every component was pluggable, you could try three transcription providers in an afternoon and find out which handled your callers' accents. That was real value in 2024, and it is the value that has decayed hardest, because the providers have all changed their interfaces since.
Learning how voice agents actually work. Reading 3,781 stars' worth of working code that handles turn detection, interruption and streaming audio is still a good way to understand the problem, and MIT licensing means you can lift whatever you find useful. As a teaching artefact it has aged far better than as a dependency.
What it is genuinely bad at, now more than ever
Anything a customer depends on. A dormant library underneath a business phone line means that when a provider changes an SDK, the line stops working and there is no upstream to fix it. That was a manageable risk in 2024. After twenty-one months without a commit it is simply the wrong foundation for a revenue-bearing process, and we would not recommend starting one on it.
It is also bad at anything a non-developer needs to see. No dashboard, no transcript browser, no analytics. If the person who cares about call quality is an operations manager rather than an engineer, there is nothing for them to look at unless somebody builds it.
Bottom line: Vocode remains a useful codebase to read and an unwise one to depend on: use it to learn how turn detection and interruption handling work, and build anything customer-facing on a maintained platform.
What Vocode actually costs - and why free is the wrong question
The library is MIT licensed and costs nothing. Your actual bill is the four providers you assemble behind it - speech-to-text, language model, text-to-speech and telephony - each billed directly by its own vendor at its own rate, plus a server to run on. On our 1,500-minute reference workload that realistically lands somewhere between $60 and $150 a month, driven almost entirely by which model and which voice you choose. A hosted offering has existed alongside the open-source core, but we could not verify its current pricing from a publicly readable page as of August 2026, so we are not quoting figures for it.
That $60 to $150 makes Vocode look like the cheapest option in the voice category, and on paper it is. The number is also close to meaningless on its own, because it excludes the only cost that actually decides this question: engineering time. Building a voice agent that handles interruptions gracefully, tolerates a slow provider, recovers from a dropped connection and does not irritate callers is genuinely difficult work - weeks rather than days for a first production-quality version, and then ongoing maintenance forever.
Set that against a bundled platform at $200 to $400 a month and the arithmetic is unforgiving. Even valuing developer time conservatively, three weeks of build is worth many months of a managed platform's subscription, and the managed platform comes with a dashboard, transcripts, monitoring and somebody to email. For a small business, the saving on paper is almost always a loss in practice.
And in Vocode's case specifically, the maintenance side of that equation has got considerably worse - which is the subject of the next section and, honestly, the reason to read this page at all.
| Layer | Published rate | Monthly at 4,000 min |
|---|---|---|
| Library licence | MIT, free | $0.00 |
| Speech-to-text | Deepgram Nova-3 streaming, $0.0048/min | $19.20 |
| Text-to-speech | Deepgram Aura-2, $0.030 per 1k characters | ~$54.00 |
| Language model | GPT-5 mini, $0.25 in / $2.00 out per 1M tokens | ~$8.00 |
| Telephony | Twilio US local inbound, $0.0085/min plus $1.15/number | $35.15 |
| Server | A small always-on instance | ~$20.00 |
| Total software cost | Four provider bills plus hosting | ~$136 |
| Engineering time | Not optional, not included, and the number that decides this | Weeks, then ongoing |
The provider stack priced at the category reference workload
At the directory's voice reference workload of 1,000 inbound calls a month averaging four minutes, or 4,000 billable minutes, a self-hosted Vocode deployment on inexpensive providers comes to roughly $130 a month. The table below prices it from rates published on 2026-08-06, using Deepgram Nova-3 for transcription at $0.0048 a minute, Deepgram Aura-2 for speech at $0.030 per 1,000 characters, GPT-5 mini at $0.25 and $2.00 per million LLM tokens, and Twilio US inbound at $0.0085 a minute.
That figure is real and it is also the least important number on this page. Add one developer week to get it working against 2026 provider SDKs, and the same week again each time something upstream changes, and the true cost of ownership is several times any managed platform's subscription. Work it through on the build vs buy comparison before you commit a fortnight to it.
Bottom line: Self-hosted Vocode runs at roughly $136 a month for 4,000 minutes of calls, and the engineering time to keep an unmaintained library working against current provider SDKs costs more than any managed platform on this list.
The real finding: the project is dormant
We check maintenance signals from the GitHub API rather than from a project's own marketing, and for Vocode the result is stark enough to change the recommendation entirely. Re-verified on 2026-08-06, the main vocode-core repository was last pushed to on 15 November 2024. Its most recent tagged release, v0.1.114a2, was published in August 2024. Widening the check to the whole vocodedev organisation does not help: no repository in it has been pushed to since November 2024, and several supporting repositories - the React SDK, the demo apps - are older still or archived.
That is approximately twenty-one months of silence, in the single fastest-moving area of applied AI. Over that period model APIs have changed, new speech and transcription providers have become standard, and the practices around latency and interruption handling have moved on substantially. A voice library frozen in late 2024 has not tracked any of it.
One further figure from the same check is worth reading carefully. The repository shows 3,781 stars and just 2 open issues. On a project of that size a near-empty issue tracker is not a sign of a clean backlog, it is a sign that people have stopped filing, because filing has stopped producing a response. Stars are a historical popularity measure that never decreases; open issues on a dead project drift towards zero. Read them together and the picture is unambiguous.
We want to be precise about what this does and does not mean. It does not mean the code stops working - MIT-licensed source on your own machine cannot be revoked, and 3,781 stars' worth of example code remains a useful thing to read. It does mean that nobody is fixing the breakage when a provider changes its SDK, nobody is reviewing a security issue, and nobody is answering the question you get stuck on. The practical consequence is that you should budget to fix compatibility problems before your first call connects, and to own every subsequent one yourself, permanently.
Our recommendation is therefore straightforward: do not start a new business-critical voice build on Vocode in 2026. If you are already running it and it works, MIT means you are not being forced anywhere - but treat the codebase as yours to maintain, pin your provider SDK versions, and plan a migration path rather than assuming upstream will come back.
Bottom line: Verified against the GitHub API on 2026-08-06: vocode-core was last pushed on 15 November 2024, has 3,781 stars and only 2 open issues, which together describe a project people have stopped filing bugs against rather than one that has fixed them.
What breaks in production, and how you find out
With an unmaintained library, the first thing that breaks is the build, and it breaks before you get anywhere near a customer. Provider SDKs move; a library pinned to the 2024 versions of several of them will need dependency work simply to install and run against current APIs. Budget for that up front and treat a successful test call as the end of phase one rather than the beginning.
Once running, the failure modes are the ones inherent to assembling four real-time providers yourself, and they are the hardest part of voice work. Latency is cumulative across transcription, model and synthesis, and past roughly a second of round trip the conversation stops feeling natural - callers start talking over the agent and the interaction degrades into the thing everyone hates about automated phone systems. Tuning that across four vendors, each with its own variability, is genuinely difficult engineering and it is precisely what you are choosing to take on.
Interruption handling - letting a caller cut the agent off mid-sentence and having it stop, listen and respond appropriately - is the single feature that separates a usable voice agent from an irritating one, and it is hard to get right. Managed platforms have spent years refining it. In a library you own that refinement.
There is no dashboard, no call history, no transcript browser and no alerting unless you build them. That means the ordinary question of what your agent said to a customer last Tuesday has no answer unless you designed a logging path in advance. Do that first, not last - and write transcripts and outcomes into a system you own as calls happen.
Finally, there is no packaged compliance story. If you need HIPAA handling or a defensible data-retention position, you design it, implement it and evidence it yourself, with no vendor attestation to point at.
If you are already running it, do these four things
Pin and vendor every dependency, so that a provider publishing a breaking SDK release cannot take your phone line down without your involvement. Second, fork the repository into your own version control and treat it as internal code rather than as an upstream dependency, which is exactly what the MIT licence exists to permit. Third, put a monitored health check on the call path that places a synthetic call and asserts on the transcript, because with no vendor status page a silent failure can run for days. Fourth, start exporting transcripts and call outcomes now, so that when you do migrate you have a test set to evaluate the replacement against.
None of that is difficult. All of it is work that a maintained platform would be doing on your behalf, which is the point.
Bottom line: On a dormant library the first failure is the build itself, so budget a developer week to get Vocode installing against 2026 provider SDKs before you count a single connected call as progress.
What it connects to that you already run
As a Python library, Vocode's integration surface is whatever Python can reach, which is everything. There is no connector catalogue and no gating: your CRM, your booking system, your database and your internal APIs are all reachable because you are writing the code that reaches them. For a team with an existing Python application, dropping voice capability into it as a library rather than bolting on an external service is a genuinely clean architecture and the main reason to consider Vocode at all.
The provider side is deliberately modular - transcription, model, speech synthesis and telephony are each pluggable and each your own account. That means the same per-component cost tuning available on the unbundled commercial platforms is available here, with the difference that you implement the plumbing. It also deploys beyond telephony, to video meetings and other channels, which is broader than most voice-only products.
The caveat that now dominates all of this: whatever provider integrations shipped in the last release are the ones you have, and any provider that has changed its interface since 2024 is your problem to adapt. The modularity that makes the library attractive in principle is only as useful as the modules, and the modules stopped being updated.
There is one integration pattern that survives the dormancy well, and it is worth naming because it is also the migration path. If the agent's only job is to hold the conversation and then fire a webhook into n8n, Make.com or Zapier with structured data, then almost none of your business logic lives in Vocode. Bookings, confirmations and CRM writes all sit in a workflow you can point at a different voice platform in an afternoon. Teams that built this way have a trivial migration; teams that put the business logic inside the agent do not.
Bottom line: Keep the agent thin and fire a webhook into an automation platform for every side effect, and replacing a dormant voice library becomes an afternoon's work rather than a rebuild.
Vocode alternatives: what to pick instead, and when
This is the most useful section on this page, because for almost every reader the correct action is to pick something else. Vocode's appeal was an MIT-licensed library with full component control and no vendor able to reprice you. Two of those three properties are available today from projects that are actually maintained.
Pick LiveKit Agents if what you valued about Vocode was the open source. It is the direct replacement and the comparison is not close: Apache 2.0, an open-source media server as well as an open-source framework, self-hostable end to end, and a repository that was pushed to on the day we checked it. It also has a free tier that covers a small business's entire agent minute volume. If you are migrating off Vocode and you still want to own the stack, this is the answer.
Pick Vapi if what you valued was component control rather than the source code. At $0.05 a minute for orchestration with transcription, model and speech passed through at cost, or free with your own API keys, you keep every provider choice Vocode gave you and hand over the latency and interruption work that consumed your engineering time. Pick Retell AI for the same reasons with an even more granular published price list. Pick Bland AI if the whole experience has left you wanting one flat rate and no components at all, and pick ElevenLabs Agents if voice quality is what your callers will judge.
The one case for staying on Vocode: an internal tool, running today, that nobody outside the business calls, where you are content to be the maintainer. Fork it, pin it, and stop treating upstream as a supplier. That is a legitimate position and MIT is exactly what makes it available to you.
| Platform | Billing unit | Cost at 4,000 min | Who it suits | Does one thing better |
|---|---|---|---|---|
| Vocode (self-hosted) | Free MIT library, you pay your own four providers | ~$136, plus the engineering time to revive it | Internal tools where you accept being the maintainer | Embeds inside an existing Python application as an import rather than an external service |
| LiveKit Agents | Agent session minutes plus your own STT, LLM, TTS and telephony | ~$166 all-in on the $50 Ship plan | Anyone migrating off Vocode who still wants to own the stack | Actually maintained: Apache 2.0, open-source media server, last pushed the day we checked |
| Vapi | $0.05/min orchestration, components at cost or free with your own keys | ~$340 including components and telephony | Teams who wanted component control, not source code | Keeps every provider choice Vocode offered while handling the latency work for you |
| Retell AI | $0.055/min base plus per-component voice, model and telephony | ~$350 on budget components, up to $1,240 on flagship ones | Teams that want to price every layer before committing | Publishes a per-component rate card, so cheap and expensive configurations are both visible up front |
| Bland AI | One flat per-minute rate, model and voices included | $560 on Start at $0.14/min, no platform fee | Businesses that never want to think about a component again | The simplest bill in voice AI: one rate covers transcription, model and premium voices |
Bottom line: If you valued Vocode's open source, migrate to LiveKit Agents; if you valued its component control, migrate to Vapi; stay only for an internal tool where you are content to fork the library and become its maintainer.
Migration and lock-in: can you get your work out?
On the pure lock-in question, Vocode scores about as well as anything can. MIT is the most permissive licence in common use: you can run, modify, fork and commercialise the code without asking anyone, and nobody can change those terms retroactively on a copy you already hold. Your model, voice, transcription and telephony providers are your own accounts. There is no contract, no renewal date and no vendor with the power to reprice you.
The irony worth naming is that dormancy inverts what lock-in means here. You are not locked in to a vendor - you are locked in to maintaining a codebase alone. That is a different risk and, for a small business, often a worse one. A vendor that might raise prices is a manageable commercial problem; a library nobody maintains, underpinning the phone line customers call, is an operational one that lands on whoever happens to be available when it breaks.
If you are migrating away, the good news is that the valuable asset is portable. Your conversation design, your prompts, your handling of edge cases and your business logic are all your own code and your own documented decisions. Moving to a managed platform means re-expressing that design in the new vendor's configuration, which is a rebuild of the plumbing rather than of the thinking.
If you are staying, fork it properly rather than depending on upstream. Pin every dependency, take ownership of the repository in your own version control, and document what you changed. That converts an abandoned dependency into an internal library, which is a supportable position - and it is what MIT exists to allow.
Bottom line: Vocode has no vendor lock-in and a severe maintenance lock-in, which for most businesses is the worse of the two: you cannot be repriced, but you are the only person who can fix it.
Vocode FAQ
What does Vocode cost?
The library is MIT licensed and free. Your actual costs are the providers you assemble behind it - speech-to-text, language model, text-to-speech and telephony - each billed directly by its vendor, plus a server to run on. On our 1,500-minute reference workload that realistically lands somewhere between $60 and $150 a month, and about $136 at the category reference workload of 4,000 minutes, depending almost entirely on which model and voice you pick. A hosted offering exists but we could not verify its current pricing from a public page as of August 2026.
Is Vocode still maintained?
No. Checked directly against the GitHub API on 6 August 2026, vocode-core was last pushed to on 15 November 2024, and its most recent tagged release, v0.1.114a2, is from August 2024. No repository in the vocodedev organisation has been pushed to since November 2024. The repository shows 3,781 stars and only 2 open issues, which on a project that size means people have stopped filing rather than that the backlog was cleared. That is roughly twenty-one months of silence in the fastest-moving area of applied AI.
Is free actually cheaper here?
Only if engineering time is genuinely free to you, which it is not. The software costs nothing, and building a voice agent that handles interruptions, tolerates a slow provider and does not irritate callers is real work - typically weeks rather than days for a first production-quality version, plus ongoing maintenance. With a dormant library you also pay a revival cost before the first call connects, because the pinned provider SDKs are two years old. Against a bundled platform at $200 a month, the break-even is a long way out.
What breaks when you build voice from a library?
Latency first: a caller notices delay above about half a second, and that budget has to cover four providers plus network time. Second, failure handling - when a provider is slow or errors, a naive implementation goes silent, which is worse than a bad answer. Third, interruption handling, which is subtle and is what separates agents that feel natural from ones people hang up on. None of these are hard to describe and all of them take real time to get right.
Who should not use Vocode?
Any business without Python engineering capacity, and any business where the phone agent needs to be reliable from day one. This is a foundation for people building something, not a route to answering your calls next week. If your goal is to stop missing bookings, a bundled platform gets you there in days and the licence cost is irrelevant next to the time difference.
We already run Vocode in production. What should we do?
Stabilise first, then plan a move. Pin and vendor every dependency so a provider's breaking SDK release cannot take your line down, fork the repository into your own version control and treat it as internal code, and put a synthetic call on a health check that asserts on the transcript, because there is no vendor status page. Then start exporting transcripts and call outcomes now, so that when you migrate you have a real test set to evaluate the replacement against.
How locked in are we?
Essentially not at all in the commercial sense, which is the strongest argument for it. MIT licensing means you can keep and modify the code indefinitely, your provider accounts are your own, and your agent logic is Python you wrote. There is no vendor who can raise your price or change your terms. The catch is that dormancy inverts what lock-in means: you are not tied to a supplier, you are tied to maintaining a codebase by yourself, which for a small business is often the worse position.
Vocode or LiveKit Agents?
LiveKit, clearly, and it is not a close call in 2026. LiveKit gives you open-source components under Apache 2.0, including the media server, with a well-supported commercial layer and a free tier that covers small-business volume - and its repository was pushed to on the day we checked. Vocode's only remaining edge is that it embeds inside an existing Python application as an import rather than an external service. If that is not your specific situation, LiveKit is the direct replacement.
What does a self-hosted voice agent really cost to run?
Add four numbers: transcription, model, speech synthesis and telephony, each per minute, then add a small server. On Deepgram Nova-3 and Aura-2 with GPT-5 mini and Twilio inbound, that comes to roughly three cents a minute, or about $136 a month at 4,000 minutes. On a flagship model and a premium voice it can be three times that. Then add your own time, which is the number people leave out and which usually dominates in the first few months.
Can you build and maintain a voice agent for us?
Yes, and with open-source foundations the handover matters more than usual, because the code is only as maintainable as its documentation. We scope it on a free automation audit, tell you honestly whether a library-based build or a bundled platform is the better economics for your volume, and if we build it you own the code, the provider accounts and the documentation outright, plus a recurring plain-language report of what ran and what needed attention.
Looking at alternatives? Vocode's lock-in position is about as strong as it gets - MIT licensed, your code, your providers, nobody can reprice you. The problem is not capability or terms, it is maintenance: our check of the GitHub API on 6 August 2026 found no activity in the project since November 2024, and an unmaintained voice library underpinning a business phone line is a risk most small businesses should not accept. If nobody in the business writes Python, this is not a realistic option at any price, including free. See the Autonoly review.
