Developer Tools6 min

Pipecat in 2026: from 1.0 to 1.15 in six months, and when to pick it over an all-in-one voice platform

Pipecat went from 1.0 in April to 1.15 in October with evals, classifiers, a multi-agent bus and expressive conversation. A review for operators deciding whether to build a voice agent on it or rent one from Vapi, Bland or Synthflow.

Erhan Timur25 October 2026Founder, Digital by Default

Share:X LinkedIn

Pipecat in 2026: from 1.0 to 1.15 in six months, and when to pick it over an all-in-one voice platform

Pipecat is the open-source Python framework for real-time voice and multimodal agents, maintained by Daily under a BSD-2 licence. It hit 1.0 on 14 April 2026 and reached 1.15.0 on 29 October, shipping a release every two to three weeks in between. If you are deciding whether to build a phone or web voice agent yourself rather than rent one from Vapi, Bland or Synthflow, Pipecat is the framework you will be building on more often than not.

This is a review for operators and the engineers they will ask to estimate the job.

What Pipecat is

A pipeline. Audio comes in over a transport (WebRTC, WebSockets, telephony, and since July Media over QUIC), passes through voice activity detection, speech-to-text, a language model with tools, text-to-speech, and goes back out. Pipecat's job is the plumbing between those services, the turn-taking, the interruptions, and the frames that carry everything. It is vendor-neutral: swap Deepgram for Speechmatics, Grok for Gemini Live, ElevenLabs for Cartesia, with a service class rather than a rewrite.

Hosting is your choice. Run it yourself, or use Pipecat Cloud, which went generally available in January 2026 at $0.01 to $0.03 per agent-minute plus $0.018 for PSTN.

What landed between 1.0 and 1.15

Version	Date	What it added
1.3.0	29 May	Multi-agent framework: every pipeline worker is a peer on a shared message bus
1.4.0	17 Jun	Pipecat Evals, a built-in behavioural test harness for agents
1.6.0	21 Jul	Media over QUIC transport alongside WebRTC and WebSockets
1.11.0	18 Sep	Evals that judge function-call arguments; Gemini Live 3.8 support; video output for LiveKit transport
1.12.0	26 Sep	Classifiers (yes/no, choice, score) for routing and quality checks; IPA pronunciation control; per-frame interruptibility
1.13.0	19 Oct	File and image send from clients; LLMWithBackend (a fast conversational model in front of a deeper reasoning worker); Grok reasoning-effort settings
1.14.0	24 Oct	New LLM and STT services (Perplexity, NVIDIA, Zoom); diarisation for NVIDIA STT; a few default-model changes
1.15.0	29 Oct	Expressive conversation (alpha): the model writes expression tags, the TTS performs them; a tone reader for emotion and likely sarcasm; an audio-quality observer; classifier-based turn completion

Two of those matter more than the rest for a business buyer. Evals means you can regression-test a voice agent the way you test software, which almost nobody buying an all-in-one platform can do today. Classifiers mean you can put a cheap, fast check ("did the caller give a postcode?", "is this an emergency?") into the pipeline without a round trip to the main model.

Adoption, roughly

Estimates vary by source: around 14,000 to 16,000 GitHub stars and roughly 300,000 weekly PyPI downloads. The precise figures matter less than the trend, which is up and steady, and the cadence, which has not slipped.

Who should care, who should wait

Build on Pipecat if

You have at least one engineer comfortable with Python and async code, and you want to own the agent rather than rent it
You need a specific combination of vendors (say, Speechmatics for UK accents, Grok Voice or Gemini Live for the model, a particular TTS voice) that no platform bundles
You need evals and classifiers in the loop for a regulated or high-volume use case
Cost per minute at scale matters and you can run Pipecat Cloud or your own infrastructure

Rent an all-in-one platform instead if

Nobody on the team will maintain Python code in six months
The job is a standard front-desk agent (answer, book, escalate) and a platform already does it well; see Voice AI for the front door
You need telephony numbers, compliance paperwork and a dashboard on day one and have no appetite for assembling them

What a Pipecat build actually involves

Transport. For phone calls, you will pair it with a carrier such as Twilio (ConversationRelay or media streams) or Daily's own telephony. See our Twilio note
Services. Pick STT, LLM and TTS. Pipecat ships service classes for most major vendors, including Grok, OpenAI, Gemini, AWS Bedrock, Deepgram, ElevenLabs and Cartesia
Turn-taking. Tune VAD and the user-turn stop strategy. 1.15's classifier-based turn completion is worth testing against the defaults
Evals. Write scenario files for your top ten call types before launch, not after the first complaint
Observability. Use the observers (audio quality, tone) and ship transcripts and function-call logs somewhere you can search
Hosting. Pipecat Cloud if you want a managed runtime; containers if you have a platform team

Pipecat vs the field

Option	Model	Best for
Pipecat	Open-source framework, optional cloud	Teams that want control and vendor choice
LiveKit Agents	Open-source framework tied to LiveKit infrastructure	Teams already on LiveKit for real-time media
OpenAI Agents SDK (voice)	Vendor SDK	OpenAI-only stacks
Vapi, Bland, Synthflow, Retell	Hosted platforms	Fast launch, less control; see our Bland and Synthflow reviews

Bottom line

Pipecat has done in six months what most voice frameworks take two years to do: a stable 1.0, a release cadence you can plan around, and the two features (evals and classifiers) that separate a demo from a system you can operate. It is still a build, and builds need owners. If you have one, Pipecat is the framework to estimate against. If you do not, read build or buy a phone agent before you commit either way.