Skip to main content

​
General

agtOS is an open-source, voice-native AI agent platform. It connects voice input/output with LLM reasoning, tool execution, and persistent memory. You can talk to it, text it, or connect it to other AI tools via MCP.It runs locally on your machine — your data stays on your hardware unless you explicitly use a cloud AI provider like Claude.
Yes. agtOS is open-source under the FSL-1.1 license — free to use, self-host, and integrate via APIs.You may need API keys for cloud AI providers (Claude, OpenAI), which have their own pricing. But agtOS itself is free, and you can run it entirely locally with Ollama at no cost.
No. A Claude API key enables cloud AI for complex reasoning, but agtOS works without one:
  • With Ollama only — local AI for chat, intent classification, and simple tasks
  • With Claude — cloud AI for complex reasoning, coding, analysis, and multi-step agent tasks
  • With both — the model router automatically sends simple queries to Ollama (free, fast) and complex queries to Claude (powerful)
Most users get the best experience with both, but you can start with either.
Two ways to pay for Claude:
  • API key (sk-ant-api03-...) — pay-per-token from console.anthropic.com. You pay only for what you use.
  • Max subscription (sk-ant-oat01-...) — flat monthly rate via your Claude subscription. agtOS uses the claude CLI as a subprocess for auth.
Both work identically in agtOS. The setup wizard auto-detects which type you have based on the token prefix.

​
Installation

  • Desktop app — best for most users. No terminal needed, includes a visual onboarding wizard, system tray with push-to-talk, and auto-updates. Everything is bundled.
  • CLI — best for developers who want source access, customization, Docker deployment, or integration into their development workflow.
Both run the same agtOS backend. The desktop app is just a native shell (Tauri 2) around it.
Partially. With Ollama and sherpa-onnx models installed locally:
  • Text chat — works offline via Ollama
  • Voice — works offline via sherpa-onnx (in-process STT/TTS/VAD)
  • Complex reasoning — needs Claude (cloud) for multi-step agent tasks
Redis is needed for cross-session memory and scheduling, but it runs locally too.
Desktop app: macOS 10.15+, Windows 10+, or Linux (x86_64). ~500MB disk for the app + voice models.CLI: Node.js 22+. Optional: Docker (for Redis), Ollama (for local AI).Hardware: Any modern machine works. Speech processing uses 2-4 CPU threads. GPU is not required but accelerates speech if available (CUDA on Linux, CoreML on macOS).
Node.js 22 provides Single Executable Application (SEA) support (used by the desktop app sidecar), stable WebSocket APIs, and performance improvements in the V8 engine that benefit real-time audio processing. Earlier versions lack these features.

​
Voice

agtOS uses a cascade pipeline:
  1. VAD (Silero) detects when you’re speaking
  2. STT (sherpa-onnx Moonshine/SenseVoice) transcribes your speech to text
  3. LLM (Claude or Ollama) generates a response, potentially using tools
  4. TTS (sherpa-onnx Kokoro) synthesizes the response as speech
All speech processing runs in-process by default — no external server needed. Total latency is approximately 500ms.
STT supports English (default), Chinese, Japanese, and Korean via the SenseVoice model. TTS supports English with 11 voice options (American and British accents). Additional languages and voices are available through the speaches fallback server.To switch STT language: SHERPA_STT_MODEL=sensevoice-int8 and SHERPA_STT_LANGUAGE=zh.
Yes. agtOS includes 11 Kokoro TTS voices. Change via the Settings page in the dashboard, or set SHERPA_TTS_VOICE in your config:
VoiceDescription
af_heartWarm American female (default)
am_adamConfident American male
bf_emmaPolished British female
bm_georgeAuthoritative British male
See Provider Configuration for the full list.

​
Memory

Three tiers:
  • Working memory — current conversation context (always available, in-process)
  • Episodic memory — past conversation summaries stored in Redis (cross-session, 30-day TTL)
  • Semantic memory — long-term facts and preferences via vector search (Redis + Ollama embeddings)
Without Redis, only working memory is available. Without Ollama, semantic search falls back to keyword matching.
Yes. agtOS can import conversation history from Claude Code, Cursor, Windsurf, Aider, and GitHub Copilot:
npx agtos memory import
Or use the Memory page in the dashboard.

​
Integration

Add agtOS as an MCP server in your Claude Desktop config file (claude_desktop_config.json):
{
  "mcpServers": {
    "agtos": { "url": "http://localhost:4100/mcp" }
  }
}
Restart Claude Desktop. You can now use agtOS tools (voice, scheduling, memory) directly from Claude.
Yes. agtOS is both an MCP server (exposes tools to external clients) and an MCP client (connects to external MCP servers). This means you can:
  • Connect agtOS to smart home servers, knowledge bases, or any MCP-enabled service
  • Connect Claude Desktop, Cursor, or custom agents to agtOS
See MCP Integration for details.
Yes. Three protocol interfaces:
  • REST API (port 4102) — 40+ endpoints for chat, memory, scheduling, devices, and configuration
  • WebSocket (port 3000) — real-time voice audio streaming
  • MCP (port 4100) — 10 built-in tools via Streamable HTTP
See API Reference for complete documentation.

​
Data & Privacy

All data stays on your machine:
  • Config — ~/.agtos/config.json and .env.local
  • Voice models — ~/.agtos/models/ (~460MB)
  • Memory — Redis (if running), stored on localhost
  • Credentials — encrypted with AES-256-GCM in the config file
Cloud AI providers (Claude, OpenAI) receive your messages when you use them, subject to their privacy policies. Use Ollama for fully local operation.
Yes. Individual conclusions can be deleted via DELETE /api/memory/conclusions/:id. Episodic memories expire automatically after a configurable retention period (default: 30 days). You can also disable memory persistence entirely via user preferences.