Open · Model-neutral · Self-hosted
Continuous voice agents
without the lock-in.
An open runtime that separates real-time interaction from model cognition. Run any model. Own every bit of the stack.
Capabilities
Built for real conversation.
Low-latency duplex, real tools, multi-agent systems, and full model freedom — without forcing everything through a text chat abstraction.
Two-plane architecture
Continuity plane handles audio, barge-in, and playout under hard deadlines. Cognition plane handles reasoning, tools, and multi-agent work asynchronously. The two never block each other.
See how it works →Continuity Plane
Capture · VAD · Barge-in · Duck · Lease · Playout
Cognition Plane
ASR · Reason · Tools · Agents · Memory · TTS
Native duplex
Preserves the native capabilities of realtime providers instead of collapsing them into turn-based chat. Cascade, hybrid, or full duplex — your choice.
Real tools & agents
Built-in tools, path-safe sandbox, multi-agent pools, research workers, and MCP support. Agents that can act, not just talk.
Client audio intelligence
Noise suppression, adaptive VAD, echo cancellation, and semantic endpointing run in the browser for cleaner, faster decisions.
Open speech stack
First-class support for open neural TTS/ASR and any OpenAI-compatible endpoint. Fully local, hybrid, or cloud — you decide.
Architecture
Designed to never wait.
Most voice systems make interaction quality wait for the model. OpenLive does the opposite.
Local-first reaction
The browser estimates speech confidence and ducks audio immediately. Server confirmation arrives later. Interaction feels instant.
Answer leases
Every generation carries a deterministic lease. New user speech invalidates the old lease. No stale audio. No race conditions.
Model neutrality
Providers are capabilities, not destinations. Swap between open models, commercial realtime endpoints, or hybrid stacks without rewriting the product.
Deterministic core
Session engine, media timestamps, and decision IDs are designed for exact replay. What happens in production can be reproduced offline.
Get started
Running in minutes.
Pure Rust. No account required. Start offline or connect any speech stack you already trust.
git clone https://github.com/byte271/Openlive.git && cd Openlive
# Offline mock — full UI, barge-in, tools, zero dependencies
cargo run -p openlive-gateway --release
# Or connect any OpenAI-compatible ASR + LLM + TTS
cargo run -p openlive-gateway --release -- \
--provider openai-compatible \
--model-base-url http://127.0.0.1:8000/v1
Then open http://127.0.0.1:8787 · Rust 1.83+ · Modern browser with mic
Why OpenLive
Closed systems have limits.
This one doesn’t.
| Closed Voice Modes | OpenLive | |
|---|---|---|
| Model choice | One provider | Any model, any provider |
| Hosting | Cloud only | Self-hosted or hybrid |
| Data ownership | Vendor terms | Yours completely |
| Tools & agents | Limited surface | Sandbox + multi-agent + MCP |
| Speech stack | Proprietary | Open neural + compatible |
| Architecture | Black box | Two-plane, leases, replayable |
| License | Proprietary | Apache-2.0 |
Principles
What we refuse to compromise.
Model neutrality
No vendor is special. Every provider is a capability you can replace.
Continuity first
Interaction quality is never allowed to wait for the model.
Real agency
Voice agents that can only talk are demos. Tools and memory are first-class.
Own the stack
Self-host, audit, extend, or replace any layer. Open is the default.
Start building continuous
voice agents today.
Clone it. Point it at any model stack. Keep complete control.