The future of agenticvoice experiences.Now.
SmartChats is an open source voice-native AI platform.
Voice in. Code, charts, web answers, and persistent knowledge — out. SmartChats turns conversation into a computing surface that gets smarter every time you use it.
One sentence in. Code, charts, memory, voice — out.
No toolbars, no copy-paste between apps. You speak. SmartChats decides which capability to invoke, runs it, shows the result, and speaks the answer back.
recorded session · loops automatically · or try it yourself →
Not a chatbot. A computing substrate.
Voice-native, not bolted on
Streaming STT → LLM → TTS pipeline designed end-to-end for real-time conversation. The whole stack is voice-aware.
ONNX VAD on-device (Silero v5) · streaming TTS · 2-chunk early-split for low first-utterance delay
Sandboxed JavaScript execution
The agent doesn't suggest code — it runs it. In a fully isolated iframe sandbox. Output renders alongside the conversation.
iframe sandbox with proxy membrane · function-call tracking · variable inspection · execution history time-travel
Persistent semantic memory
Every conversation builds an entity-relation knowledge graph. Embeddings make it searchable. Future conversations are smarter.
SurrealDB triple store · HNSW vector index · 1536-dim embeddings · multi-hop relation traversal
Multi-provider, no lock-in
Switch between GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro mid-conversation. Or bring your own keys and route directly.
Unified streaming protocol across providers · structured-output coercion · BYO-key support on every plan
MCP-native interoperability
First-class Model Context Protocol support — bidirectional data export and ingestion, third-party tool wiring, and integrations across the agent ecosystem. Your data stays portable; your platform stays composable.
bundled smartchats-mcp server · cloud → local round-trip data portability · pluggable into any MCP client (Claude, Cursor, IDEs) · MCP tools surface as agent functions automatically
An open-core stack you could read end-to-end.
Voice → orchestration → multi-modal output. Three layers, clearly separated, swappable. The full stack — frontend, local server, schema, voice pipeline — ships open-source under MIT, so the entire app can be self-hosted. The production cloud backend (billing, auth, multi-tenant) stays private, and provides data synchronization across all devices for subscribing users.
Voice Layer
Tivi- ·ONNX VAD (Silero v5)
- ·Streaming STT
- ·Per-utterance TTS
- ·Optimized latency / cost that scales to millions of simultaneous users
Cortex Engine
Open core- ·Multi-provider LLM router
- ·JSON-stream parser
- ·Function-calling loop
- ·Background processes
Output Surface
Multi-modal- ·Streaming TTS audio
- ·Sandboxed JS exec
- ·Knowledge graph writes
- ·Visualization widgets
From stealth to platform.
A focused four-quarter path: ship the open core, layer the hosted product, build out the integration ecosystem, then turn the platform into something other apps can embed.
Stealth
Working prototype with full stack operational — voice pipeline, knowledge graph, sandboxed execution, multi-provider LLM routing, and billing infrastructure all live. Active daily development.
Open Core + Hosted
Open-source the core (MIT). Hosted web app product with managed billing, auth, and infrastructure ships publicly to prosumer users.
Integrations + Mobile App
SmartChats Mobile App launch. Third-party integrations buildout — Gmail, Calendar, X, GitHub, and more — so SmartChats can read, write, and act across the apps where users already live.
Enterprise
Drop-in voice agent SDK for enterprises — a JS / web component lets any external product embed SmartChats' full stack (voice, KG, sandboxed execution) without building it themselves. White-label and self-hosted options for regulated environments.
Three forcing functions converged in 2025.
Frontier models cleared the conversational quality threshold
GPT-5, Claude Sonnet 4.5, and Gemini 2.5 are the first generation that can reliably handle multi-turn voice context, structured tool use, and code generation in a single coherent loop.
On-device inference made always-on voice tractable
ONNX-runtime VAD models like Silero v5 ship sub-50ms speech detection in the browser. No more push-to-talk; conversation can be ambient.
Streaming TTS at usable price points
OpenAI's gpt-4o-mini-tts and equivalents made per-utterance synthesis cheap enough to run continuously. Voice can be the primary output, not a feature.
Voice-as-interface is older than the GUI. What changed is that — for the first time — the substrate underneath voice can actually do anything. SmartChats is the bet that when those two things meet, the result is a new kind of computer.
The next great computing interface is conversational, programmable, voice-native.
The keyboard didn't replace the punchcard because it was faster — it replaced it because it matched how humans think. Voice plus programmable substrate matches how humans think now, with a knowledge layer that compounds. That's the platform we're building.
Sheun Aluko, MD, MS
Founder & CEO
Sheun is a software engineer building at the intersection of voice, AI, and personal computing. SmartChats is the convergence of years spent shipping production AI systems and a personal frustration with how much of human cognition still happens at a keyboard.
The product is built solo, end-to-end: the voice pipeline (Tivi), the agent engine (Cortex), the knowledge graph layer, the sandboxed execution environment, the multi-provider routing, the billing system. Every part is shipping.
Talk to it.
The fastest way to understand SmartChats is to say something to it. Live demos by request — partners, investors, and program reviewers get a guided walkthrough plus access to a working session.