changelog

What shipped, dated.

The platform is in developer preview and moves weekly. This is the record — the runtime, the platform, the SDKs and the demos, in the order they landed.

voqalize/voqalize Docs Quickstart
demosplatform

Ten demos, one gallery, new home for the brains

  • The demo gallery moved to the apex at /demos — ten working agents, each with its source in the public repo.
  • Flowforge joins the set: a voice-driven workflow studio that acknowledges every request on screen as it lands.
  • Demo brains now run as a container on the voice node instead of Cloud Run. One fewer network hop between the runtime and the brain it dials.
  • Legacy console demo routes redirect to the apex. Old links keep working.
runtime

Turnguard — interruption that learns the room

  • The fixed barge-in threshold is gone. Turnguard scores each incoming utterance against what the session has already taught it, so a cough or a background voice stops cutting the agent off mid-sentence.
  • Speech that arrives between end-of-turn and first playout now interrupts immediately. That collision window used to swallow the interruption entirely.
platformsdk

Hosted MCP server, React client, docs site

  • An MCP server at app.voqalize.com/mcp, hosted and OAuth-authenticated. Create agents, mint keys, point brain URLs and read call logs from your editor’s agent — nothing to install.
  • @voqalize/client-react ships useVoqalSession: mic, WebRTC, reconnects and server messages behind one hook.
  • Documentation moved to a real docs site at /docs.
platform

One brain URL. No deployment modes.

  • An agent’s brain is now a single WebSocket URL. The runtime never interprets where it points — your own inbound server, or our relay if you can’t accept inbound connections.
  • The runtime presents an identical token to every brain, so every brain verifies it the same way. Per-agent audience claims are gone.
  • Push-triggered deploys for the voice runtime, dev and production.
sdk

The Python SDK dropped pipecat

  • Brain is the whole surface: on_interaction, i.inference(), i.action(). No frame processors, no pipeline to assemble, no pipecat in your dependency tree.
  • The Go SDK verifies our signing keys by default, with the public keys embedded. Authentication is correct with zero configuration.
runtimesdk

Change the voice mid-call

  • session.configure_tts swaps voice, language or model without dropping the call.
  • session.configure_stt retunes voice activity detection and turn-taking mid-call; a model change applies at the end of the current turn rather than cutting the speaker off.
runtime

Every session on one UDP port

  • WebRTC media for all concurrent sessions multiplexes onto a single UDP port, IPv4 and IPv6. A port per session was never going to scale past one node — this is what more than one node is built on.
  • pipecat 1.2.1 → 1.5.0.

Build on it while it moves.

The SDK and the wire protocol are still changing, and we say so on every page. What isn't changing is the shape: your brain stays in your building.

Start the quickstart Talk to us
Talk to a live agentservicing desk · real session