The best open source voice AI agent frameworks in 2026

Picking a voice agent framework in 2026 means choosing what you are willing to operate. One project hands you a media server and a phone network. Another hands you three interfaces and gets out of the way. The gap between those two answers is much larger than the feature tables suggest, and it decides how much of your week goes into plumbing.

Micdrop publishes this ranking, and Micdrop appears in it at number three. That is worth knowing before you read a word of the rest. Every competitor here links to its own site, the figures come from each project’s own repository, licence file and pricing page, and the criteria below say plainly where Micdrop places last.

Twelve projects are ranked, in two separate groups. The first eight are open source frameworks you run yourself. The last four are hosted platforms, kept apart because comparing a library you install with a service you rent on the same scale produces nonsense. A framework has no price per minute. A platform has no repository to read.

Every repository, licence file, pricing page and release history below was read on 13 August 2026. Voice moves fast enough that some of it will have shifted by the time you get there.

How the twelve were judged

Seven criteria decided which projects made the list and what each entry has to report.

• The server-side language is taken from the project’s own documentation, since a client SDK in a language says nothing about what the orchestration runs on.
• What you have to operate covers the media server, the worker process or the separate Python service each project adds to your deployment.
• The licence is read as written, including the conditions some projects attach to an otherwise permissive one.
• Provider independence asks whether speech-to-text, the model and text-to-speech are swappable, and what happens when one of them goes down.
• Reach counts telephony, video, mobile and embedded targets.
• Activity and community come from stars, contributors, download volume and release cadence, reported inside each entry.
• Public production evidence means named customers confirmed on a first-party page. Several well-known projects have none, and each entry says so.

The order is a different question, and no weighted sum of those seven produces it. The eight frameworks are ordered by how wide the set of projects is that each one is the right default for. LiveKit Agents and Pipecat come first because they fit almost any voice project, whatever the language or the channel. Micdrop comes third because the shape it fits, a voice mode inside a TypeScript web application, is a very common one, and only two other projects here cover it. Below that sit the projects that are excellent inside a narrower boundary: a single model vendor for the OpenAI Agents SDK, which covers that same shape and asks you to buy the whole pipeline from OpenAI, then a restrictive licence for TEN, video for Vision Agents and telephony infrastructure for Jambonz. Vocode is last because it stopped shipping. A ranking by community size alone would run in a different order, and each entry gives you the numbers to build that one instead.

The last two criteria cut against the project publishing this article, so they deserve a paragraph here rather than a line buried further down. Micdrop is by a wide margin the smallest and youngest entry in this ranking, with a community a fraction of the size of Pipecat’s or LiveKit’s and no customer page to point at. It also ships no infrastructure: no media server, no phone numbers, no managed plan, no support desk, and a single maintainer behind the packages. Those limits apply to every use case below, not only to the ones where they are convenient to mention.

A note on what is excluded. General TypeScript agent frameworks appear in a lot of 2026 voice lists and do not belong there. Mastra’s own documentation hands the audio loop to LiveKit for voice activity detection, turn detection and barge-in, and VoltAgent’s voice package wraps speech-to-text and text-to-speech rather than running a conversation. Voiceflow is left out for the same reason: voice is a phone channel on a chat-first design canvas, with no barge-in, turn-taking or voice activity controls exposed at all.

The comparison at a glance

Open source frameworks you run yourself

1. LiveKit Agents

LiveKit is a WebRTC media server with an agent framework on top, and it is the broadest answer on this list. The framework is Apache 2.0, the Go media server is Apache 2.0, and so is the SIP gateway. The Python agents repository sits near 13,000 stars with 446 contributors and over a million PyPI downloads a week. The company raised a $100M Series C in January 2026 at a $1B valuation, led by Index Ventures.

The production evidence is the strongest in the field and it is public: OpenAI runs ChatGPT’s voice mode on it, alongside SAP, NVIDIA, Tesla, Spotify and Salesforce on LiveKit’s own customer page. Client SDKs cover JavaScript, React, Swift, Kotlin, Flutter, React Native, Unity, Rust, C++ and ESP32, which no other project here matches.

The Node distribution is real and first class in the documentation, with its own starter template alongside the Python one. It carries roughly 37 plugins against Python’s 72, and MCP support is still marked Python only in the tool documentation. The semantic turn-detection models also ship under a separate LiveKit Model License stating they may only be used together with the LiveKit Agents framework, never standalone or with another framework. Self-hosting loses that turn detector along with the noise cancellation plugin. LiveKit Cloud opens free, then $50 and $500 a month, with agent minutes at $0.01 on top.

2. Pipecat

Pipecat, from Daily, is the reference Python framework and the most starred agent framework here: 14,000 stars, 293 contributors, over 300,000 PyPI downloads a week, and a release every two or three weeks since 1.0 landed in April 2026. That 1.0 is only four months old, so the stable API is younger than the project.

Its licence is the cleanest of the two leaders. Pipecat is BSD 2-Clause, and so is its smart-turn detection model, so nothing about turn detection binds you to the framework that trained it. That single detail is a real advantage over LiveKit for anyone who cares where their models can run.

The transport catalogue is the widest of any framework on this list: ten transports, among them WebSocket and three flavours of WebRTC, plus first-party telephony serializers for Twilio, Telnyx, Plivo, Exotel, Genesys and Vonage, and SIP with call transfers. Client SDKs reach JavaScript, React, React Native, Swift, Kotlin, C++ and ESP32. Pipecat Cloud went generally available in January 2026 at $0.01 to $0.03 per agent minute, with PSTN at $0.018.

The framework is Python only, and its own README says so in the first line. Every TypeScript repository in the organisation is a client or a tool. A Node team therefore runs a second service in a second language. Pipecat publishes no named production customers, a gap that stands out given how often it is described as the industry default.

3. Micdrop

Micdrop is here because of one gap that the eleven others leave open. Adding a voice mode to a web application normally costs at least one of three things: a second service to deploy, a media plane to operate, or a per-minute fee stacked on your provider bills. Micdrop asks for none of the three. The browser calls Micdrop.start({ url }), the server hands its existing socket to new MicdropServer(socket, options), and the conversation runs inside the application you already ship, on the providers you already pay.

The MIT-licensed packages carry the microphone, voice activity detection and playback in the browser, the model, speech-to-text and text-to-speech on the server, and nothing else. Three behaviours that usually cost a sprint are options here. Provider fallback replays buffered text on a second vendor when the first one fails, semantic turn detection asks the model whether the user actually finished, and a fully European stack runs on Mistral, Gladia and Gradium without touching orchestration code. Ten packages have shipped continuously since 2025, past their 2.0.

One entry above avoids the same three costs. With the OpenAI Agents SDK, either the browser talks straight to OpenAI and your server drops out of the conversation, or you keep the server in the loop and write the audio capture and playback yourself, and either way the model, the transport and the bill come from one vendor.

Shipping products run the pipeline, Raconte for voice interviews led by an AI and Cibli for recruitment where candidates answer out loud, though Raconte is built by Micdrop’s own maintainer and neither publishes a case study. The rest of the picture runs the other way. There is no telephony, no video, no mobile SDK, no managed hosting and no customer page, the community is the smallest in this ranking by a wide margin, and Pipecat and LiveKit beat it on every criterion that involves reach, adoption or the size of the team behind the code.

4. OpenAI Agents SDK

The OpenAI Agents SDK is MIT-licensed, ships for both TypeScript and Python, and its realtime layer genuinely runs on a server. The transport guide is explicit that WebSocket is the default server choice, suited to server-side use cases, telephony bridges and custom audio pipelines, while WebRTC is the browser default. On that path you own the audio capture and playback and the SDK owns the event loop, tools, guardrails and history. The Python SDK sits above 28,000 stars, the JavaScript one near 3,600, and both ship a release most weeks.

The trade is the vendor, not the language. The realtime path is OpenAI’s speech-to-speech model rather than an assembled pipeline, so there is no separate speech-to-text and text-to-speech chain to point elsewhere. Every shipped realtime transport targets OpenAI’s own API, and the adapter that makes text agents provider-agnostic does not apply to realtime agents. Swapping providers means implementing the transport layer yourself. Telephony is covered, with a native SIP transport for phone calls and a Twilio transport still flagged as beta. Pricing is metered per token on the Realtime API, at $32 per million audio input tokens and $64 per million output, which third parties translate to roughly ten cents a minute. Both SDKs are still pre-1.0 after seventeen months.

5. TEN Framework

TEN, backed by Agora, is one of the most active projects in the field: 11,000 stars, a release roughly every two weeks, and commits landing daily. It is a polyglot graph runtime where Go, Python, C++ and Node extensions share one message bus, with its own well-regarded voice activity detection and turn-detection models, multimodal video streams and a visual designer. It is also the renamed Astra, so lists that rank both are counting one project twice.

Read the licence carefully. GitHub declines to classify it as a standard licence, because it is Apache 2.0 with additional conditions. The text forbids hosting the framework on end-user devices, mobile terminals included, and forbids deployment in a way that competes with Agora’s offerings or lets others compete with them. Since Agora sells a conversational AI engine, that second clause plausibly rules out building a voice platform business on TEN. Only the packages/ directory is unrestricted. The framework is also still on 0.x after two years, so no API stability is promised, and no named production users appear anywhere on the project’s site, README or press material.

6. Vision Agents

Vision Agents, from Stream, is a year old and already past 8,000 stars, with weekly releases and Apache 2.0 throughout. It is strongly provider-agnostic across speech-to-text, text-to-speech, models and speech-to-speech, and runs WebRTC over Stream’s edge network with a free tier of 333,000 participant minutes a month. Client SDKs cover React, Android, iOS, Flutter, React Native and Unity, and telephony arrives through Twilio or Telnyx.

What sets it apart is vision. These are agents that watch a video stream rather than only listening to one, which opens telehealth, live coaching and anything where the model needs to see what the user sees. The framework is Python only, with no Node server path, and it remains pre-1.0 with no named production users published.

7. Jambonz

Jambonz is the answer to “is there a mature Node.js voice project”, and the answer comes with a qualification. It is MIT, written in JavaScript, maintained by FirstFive8 since 2019, and still shipping in August 2026. It is also telecom infrastructure rather than a browser voice framework: SIP trunks, carrier-grade routing, media control, and swappable speech and model vendors underneath.

Its user list is unusual, because it is a list of the other entries: Retell AI, Vapi, Cognigy, Rasa and Movius all run on Jambonz. If you are building phone-first agents in Node and want to own the telecom layer, this is the project, and it complements rather than competes with anything that handles the browser leg. Star counts understate it badly, since the work is spread across about thirty repositories, and that spread is also the warning. Adopting Jambonz means running a telecom stack, with the SIP knowledge and the operational surface that implies, which is a great deal more to carry than a voice library if all you wanted was a browser talking to your backend.

8. Vocode

Vocode is MIT, Python, a Y Combinator W23 company, and still ranked fourth to sixth in most 2026 lists out of habit. It earned that reputation: 3,800 stars, a clean FastAPI telephony server with Twilio and Vonage support, and one of the first credible open source answers to voice agents.

It has not moved since November 2024. The last stable release is from June 2024, vocode.dev now redirects to the GitHub organisation, the hosted API’s subdomains have no DNS records at all, and Y Combinator’s own company page marks Vocode as acquired without naming an acquirer. The repository is not archived and the documentation carries no deprecation notice, so it still looks alive from a search result. It is ranked here so the picture is complete, and it is the clearest evidence that a permissive licence and a healthy star count say nothing about whether a project will still build next year.

Hosted platforms, for comparison

These four are rented rather than run. None of them is open source, and each ships what its documentation calls a server SDK, which in every case is a typed REST client rather than an orchestrator. The voice loop stays on their infrastructure.

Most of the seven criteria above cannot be applied to a platform with no repository to read, so these four are ordered on the one axis that stays measurable: how much of the stack you keep control of, from Vapi where you can supply your own model keys down to Synthflow where there is no self-hosting and no SDK at all.

9. Vapi

Vapi is the closest hosted analogue to the philosophy of the frameworks above, because it genuinely supports bringing your own keys: the platform fee is $0.05 a minute and model providers are billed at cost, or at zero if you supply the credentials. It raised a $50M Series B in May 2026 at around a $500M valuation, has handled over a billion calls, and publishes real logos including Amazon Ring, which routes all of its inbound calls through it. It carries SOC 2, GDPR and PCI DSS, and it is the only one of the four publishing both an SLA figure, at 99.9%, and a public status page.

The orchestration engine has no public repository, the browser transport is WebRTC through Daily, and independent analysis puts realistic all-in cost between $0.23 and $0.33 a minute. HIPAA is a $2,000 monthly add-on.

10. Retell AI

Retell AI is the capital-efficiency story of the group: a $4.6M seed and no Series A, against $60M of annualised revenue reported in April 2026, growing 650% year on year, and profitable. It handles more than 50 million calls a month. Its telephony reach is the broadest here, with verified SIP integration into Twilio, Telnyx, Vonage, Avaya, Genesys Cloud, Five9 and Amazon Connect, and its SDKs are Apache 2.0 even though the core is closed.

Pricing is unbundled despite the single headline range of $0.07 to $0.31 a minute, adding voice infrastructure at $0.055 a minute to text-to-speech, telephony and a per-minute model line, which lands near $0.245 for a frontier model over Twilio. Two facts matter for European buyers: Retell’s compliance documentation states plainly that it operates no services within the European Union, and all of its case studies are anonymised, with no named customer logos published.

11. Bland

Bland runs its own GPUs and its own voice, model, speech-to-text and text-to-speech stack rather than wrapping third-party APIs, which makes it the most vertically integrated option and the only one with no bring-your-own-key story at all. Pricing is fully bundled in exchange: $0.14 a minute with no platform fee, down to $0.11 on a $499 monthly plan, with speech and model usage included. On a frontier model that works out cheaper all-in than Retell.

It raised a $50M Series C in June 2026 led by Dell Technologies Capital, serves more than 250 enterprise customers including Samsara and CNO Financial Group, and is the only platform of the four documenting on-premise and VPC deployment with zero data egress. One caution on wording: Bland’s documentation describes itself as “fully self hosted”, which means Bland hosts its own hardware rather than that you can host Bland. Its published scale figures also conflict between its own site and press coverage.

12. Synthflow

Synthflow is the one hosted vendor with a genuine European story. The company is AgentFlow AI GmbH, registered in Berlin, and it publishes EU and US hosting regions, ISO 27001, SOC 2, GDPR, HIPAA, PCI DSS, a 99.99% SLA and an EU AI Act transparency statement. It runs its own carrier network with regional points of presence, integrates with a dozen named contact-centre platforms including Five9, Genesys Cloud, Talkdesk, RingCentral and Zoom, and handles over 10 million calls a month across 30 countries for customers including Freshworks and Thryv.

There is no self-hosting and no SDK. The documented surface is a REST and streaming API, and what the marketing calls a WebSocket SDK is a protocol you implement yourself from copy-paste samples. The pricing page now publishes a single number: enterprise contracts start at $30,000 a year. The component rates that blend out to roughly $0.15 to $0.24 a minute come from third-party sources that agree with each other rather than from Synthflow directly.

How to choose between them

Start from what you refuse to operate, because that eliminates most of the list in one step.

If voice is your product and it involves phone numbers, video or a dozen client platforms, LiveKit Agents or Pipecat are the two serious answers, and the choice between them comes down to language and licence. Python teams take Pipecat and get the cleaner licence with no model carve-out. Teams that want a managed media plane with observability take LiveKit and pay per minute for it.

If your backend is Node and the voice feature lives inside a web application you already ship, three projects qualify: LiveKit Agents, the OpenAI Agents SDK and Micdrop. Separate them on two questions. Does a WebRTC media plane earn its keep for a one-to-one call between a browser and your own server, and are you willing to buy the model, the transport and the bill from one vendor? LiveKit brings a much larger community, telephony and a managed option, at the cost of an SFU and a worker process. The OpenAI Agents SDK is the least code of the three and the shortest path if you have already committed to OpenAI, since swapping any part of the pipeline means writing a transport layer. Micdrop brings a WebSocket into the server you already run with three swappable provider slots, at the cost of everything the other two do beyond that. We wrote the long version of two of those comparisons separately, against Pipecat and against LiveKit Agents.

If phone calls are the whole product, look at Jambonz to own the telecom layer in Node, or at a hosted platform to skip it entirely. Regional needs change the answer again. Outside the twelve ranked here, Bolna is MIT, actively released and built around Indian-language telephony, which none of the leaders cover well.

Among the hosted four, the deciding question is where your money and your data go. Vapi is cheapest if you bring your own keys and can live with a $0.05 floor per minute. Bland is simplest to forecast because everything is included, and impossible to unbundle. Retell has the widest contact-centre reach and no European presence. Synthflow is the one to shortlist when EU hosting is a requirement and a $30,000 annual floor is acceptable.

Two questions get skipped at selection time and asked in production. What happens when your text-to-speech provider goes down in the middle of a call? What happens to the conversation when the user pauses mid-sentence? Provider fallback and turn detection both deserve a test while you still have the choice.

Frequently asked questions

Getting started

The short summary of this ranking is that Pipecat and LiveKit Agents are the two projects most teams should look at first, and that they ask you to adopt a Python service or a WebRTC media plane in exchange for what they cover. That trade is worth it for phone calls, video and scale. It is a lot of machinery for a voice mode inside a web app.

That narrower case is what Micdrop is for, in TypeScript on both sides, over a WebSocket to the server you already deploy:

npm install @micdrop/server @micdrop/client @micdrop/openai @micdrop/gladia @micdrop/elevenlabs

The Getting Started guide has a working call running in about five minutes, and the React hooks cover the UI states you will want to render.