TEN Framework
A real-time conversational AI framework from Agora with a visual designer, a SIP extension and turn detection, released under Apache-2.0 plus conditions that forbid competing deployments, which makes it source available rather than open source.
What it does
TEN "is an open-source framework for real-time multimodal conversational AI", in its own words. It ships two assistant shapes, a chained one (speech to text, language model, text to speech) and a speech-to-speech one, plus a visual designer called TMAN Designer where you pick the extensions and set keys. Extensions cover voice activity detection, turn detection ("enables full-duplex dialogue communication"), avatars from HeyGen, Tavus and Trulience, memory, and a SIP extension "that enables phone calls powered by TEN", with a Twilio example. It runs on the ESP32-S3 Korvo board as well as servers.
What it replaces
TEN replaces Vapi, Retell AI and ElevenLabs Agents for a team that can accept its licence conditions and its dependency on Agora's real-time service for the documented quick start.
The licence is the reason this page labels it source available. The LICENSE file grants Apache-2.0 "with the following additional conditions": you may not host it "on any End User devices", and you may not deploy it "in a way that competes with Agora's offerings and/or that allows others to compete with Agora's offerings". Derivative works stay under the same terms. Only the packages directory is plain Apache-2.0. A field-of-use restriction like that is not something an OSI-approved licence permits, and the README's own badge reads "Apache 2.0 with certain conditions".
Where it falls short beyond the licence: the quick start requires an Agora app ID and certificate, an OpenAI key, a Deepgram key and an ElevenLabs key. The README does not describe running without Agora's service.
How it runs
Docker and Docker Compose with Node.js 18 LTS: docker compose up -d, then task install and task run inside the container. The designer runs on port 49483. Self-hosting is described as building the example Dockerfile, or splitting a Go backend on port 8080 from a Next.js front end. GitHub Codespaces is offered as a Docker-free path. Minimum stated hardware is 2 CPU cores and 4 GB of RAM. No hosted TEN is sold in the README.
Who it fits
A team already on Agora that wants a visual designer and a SIP path. A team that needs an OSI licence or wants to run without Agora should use Pipecat or LiveKit Agents.
Limits the README states
The 2-core, 4 GB minimum; a first build of the default example takes 5 to 8 minutes by the README's estimate; running the transcriber without Docker is marked beta.
Sources
- TEN Framework README: github.com/TEN-framework/ten-framework, read 2026-09-16.
- TEN Framework LICENSE, "Open Source License" with additional conditions, copyright Agora: LICENSE, read 2026-09-16.
More voice agent frameworks
A Python framework under BSD-2-Clause, maintained by Daily, for real-time voice and multimodal agents, with the widest provider list on this site and telephony serialisers for Twilio, Telnyx, Plivo, Vonage, Exotel and Genesys.
An Apache-2.0 framework for real-time voice agents that connects any speech, language and voice provider, with semantic turn detection, native MCP tool support, a test framework, and an open-source media and SIP stack it can run on entirely.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.