DEV Community

Cover image for Pipecat or Vocode? Ask Who Patches It When a Provider Changes
Pritesh Kumar for Dograh AI

Posted on

Pipecat or Vocode? Ask Who Patches It When a Provider Changes

Here's a ten-minute test we'd run before building on either Pipecat or Vocode. Clone the repo, bump one provider client to its newest version, the OpenAI SDK is a good pick, and place a call. Whatever breaks is now your job to fix, and how often that happens depends mostly on whether anyone upstream is still shipping.

On paper the two look like cousins. Both are open-source Python frameworks for real-time voice agents, and both carry permissive licenses, BSD-2-Clause for Pipecat and MIT for Vocode. Either one can go inside a closed commercial product as long as the copyright notice travels with it, and either can be forked freely. So the license won't settle this one.

One of them stopped moving

Pipecat put out v1.16.0 on 12 November 2026, about two weeks after v1.15.0, and commits were still landing days later. Daily, the company behind the Daily video platform, maintains it with outside contributors, and new carrier serializers and services keep showing up in each release.

Vocode's last commit is dated 15 November 2024, a client update for websockets and Cartesia. The last stable release on PyPI is 0.1.113 from June 2024. The README still asks for community maintainers. The repo isn't archived, but the hosted dashboard link in that README no longer loads, and vocode.dev now forwards to the GitHub organization.

That gap hurts more in voice than in most corners of Python. Model providers change their APIs and retire models every few months. Vocode pins clients from 2024, including an OpenAI client from the 1.30 line and an early langchain-openai release. The first time a provider retires a model you depend on, the fix is yours to write and maintain. Pipecat isn't free of upkeep either. Releases come fast, so pin a version and rerun your call tests after every upgrade.

The phone call muddies the picture a bit

Here's the part that cuts the other way. On transfers, the quiet project ships more out of the box. Vocode has a TransferCall action for Twilio and Vonage. On Twilio it redirects the live call to a new number through Twilio's API, which makes it a blind transfer with no check that anyone picks up.

Pipecat reaches far more carriers, with serializers for Twilio, Telnyx, Plivo, Exotel, Vonage, Genesys and Jambonz, plus SIP and Daily's phone service. Its telephony guide is upfront that the WebSocket routes for Twilio, Telnyx, Plivo and Exotel carry no advanced call center features like transfers or reconnects. On those routes your team writes the handoff against the carrier's API, or moves the call onto SIP or Daily's phone service, where transfers are supported.

So a working Vocode app on Twilio with a blind transfer that does its job is no emergency. Pin the versions and keep a plan ready for the day a provider retires a model.

Moving between them means rewriting the agent

The two frameworks model a call in different shapes. In Pipecat a call is a pipeline of processors passing frames of audio and text along, so adding a filter or a tool call means dropping another processor into the line. In Vocode a call is a conversation object that ties a transcriber and a synthesizer to an agent, with actions hanging off the agent.

Two workshops side by side. In the Pipecat one, frames of audio and text ride a belt past a line of processors while a new filter step is lowered into the line. In the Vocode one, a single conversation object holds a transcriber, an agent and a synthesizer, with actions on the agent. A line from either workshop leads to developers rewriting the agent code.

There's no adapter between those shapes. Porting a Vocode app to Pipecat is a rewrite of the agent code, so budget it like a rebuild, with its own round of real phone calls before cutover.

The framework is the small part of a phone agent

Pick either and you get a loop that talks. A production phone agent also needs an editor that non-developers can use, outbound lists with retries, call history with recordings, QA on finished calls and a deployment that scales. Pipecat at least has a hosted runtime in Pipecat Cloud, run by Daily. Vocode has none that loads today, so you host it yourself.

Our standing view is that Pipecat is an excellent base, and production deployment still needs significant plumbing on top. That plumbing is what Dograh, the open-source AI voice agent platform, adds around a fork of Pipecat it keeps in step with upstream releases. Campaigns there dial outbound lists with retries and calling hours, and every call keeps its transcript and recording alongside a per-turn trace. Its Transfer Call works on Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI, and it calls the destination first, connecting the caller only when someone answers.

If you're starting fresh, start on Pipecat. Then run the bump-one-client test anyway, and try a transfer over a real phone call before anything ships, whichever framework you end up on.

Anyone here still running Vocode in production? What broke first when a provider changed its API?

Release dates and carrier support for both sit in one table in our rundown of Pipecat and Vocode maintenance and phone support.

Top comments (0)