Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

Appearance settings

paralizeer/tinyagentos

Open more actions menu
 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

42 Commits
42 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TinyAgentOS

Self-hosted AI agent platform for low-power hardware. Like Umbrel but for AI agents — app store, model management, agent deployment, and system monitoring in a single web dashboard.

Run AI agents on a $150 Orange Pi, a budget x86 PC, or anything in between. Flash an image, open the browser, deploy your first agent.

Quick Start

pip install -e .
python -m uvicorn tinyagentos.app:create_app --factory --host 0.0.0.0 --port 8888

Open http://your-host:8888 in your browser.

Features

App Store

Browse and install agent frameworks, LLM models, and services from the built-in catalog. Hardware-aware recommendations — the store shows what works on your hardware.

Model Manager

Download and manage LLM models with support for multiple formats (GGUF, RKLLM). Automatic variant selection based on your hardware profile (NPU, GPU, CPU).

Agent Deployment

Create agents through a wizard — pick a framework, choose a model, configure, and deploy. Each agent gets its own LXC container with isolated memory and QMD serve.

Dashboard

Real-time KPIs, backend health monitoring, agent status, and query latency metrics with Chart.js graphs. Auto-refreshes via htmx.

Memory Browser

Search and browse agent memories using QMD's full-text search. Filter by agent and collection, view chunks, delete entries.

System Config

YAML editor with validation, hardware profile display, catalog sync.

App Catalog

Type Apps
Agent Frameworks SmolAgents, PocketFlow, OpenClaw, Swarm, OpenAI Agents SDK, Langroid
Models Qwen3 Embedding 0.6B, Reranker 0.6B, 1.7B, 4B, 8B
Services Gitea, Code Server, Open WebUI, n8n, Dify

Models include RKLLM variants for Rockchip NPU and GGUF for CPU/GPU. Community can contribute apps via PR to the catalog repo.

Supported Agent Frameworks

Framework Type Notes
OpenClaw Full agent framework Multi-channel (Discord, Telegram, Slack, Signal)
nanoclaw Lightweight OpenClaw Minimal footprint
picoclaw Micro OpenClaw Smallest possible agent
SmolAgents Code-based agents 26k stars, 30% fewer LLM calls
PocketFlow 100-line framework Zero deps, graph-based, MCP support
TinyAgent Edge tool calling 1-3B models with LLMCompiler
Hermes Function calling Nous Research
Agent Zero Autonomous agent Self-correcting workflows
Swarm Multi-agent handoffs OpenAI, lightweight, experimental
OpenAI Agents SDK Multi-agent orchestration Provider-agnostic, tracing, guardrails
Langroid Message-passing agents Built-in vector store, local LLM support

Supported Backends

Backend Type API
rkllama Rockchip NPU Ollama-compatible
ollama CPU/GPU Native
llama.cpp CPU/GPU/Vulkan OpenAI-compatible
vLLM GPU (CUDA/ROCm) OpenAI-compatible

Architecture

TinyAgentOS (FastAPI + htmx, port 8888)
├── Dashboard  — KPIs, backend health, metrics charts
├── App Store  — browse/install frameworks, models, services
├── Models     — download, manage, assign LLM models
├── Memory     — search/browse agent memories
├── Agents     — deploy, manage, start/stop, logs
└── Settings   — config editor, hardware profile

Infrastructure:
├── App Registry      — manifest parsing, install tracking
├── Hardware Detector  — CPU, RAM, NPU, GPU, disk profiling
├── Installers        — pip/venv, Docker Compose, model downloads
├── Container Manager — LXC via incus (agent isolation)
├── Health Monitor    — background polling, metrics SQLite
└── Catalog Sync      — git-based app catalog updates

Agent data always lives in the agent's LXC container. TinyAgentOS accesses memory via each agent's QMD serve instance over HTTP. This enables multi-host fallback.

Service apps use Docker. Gitea, Code Server, and other services run as Docker containers managed via Docker Compose.

Hardware Targets

Profile Example Hardware Notes
ARM + NPU
arm-npu-16gb Orange Pi 5 Plus (RK3588) Primary dev target, NPU accelerated
arm-npu-32gb RK3588 32GB boards More concurrent models
arm-npu-64gb+ High-RAM ARM boards Full model suite
NVIDIA (CUDA)
x86-cuda-4gb GTX 1050 Ti Small quantized models only
x86-cuda-6gb GTX 1060 6GB, RTX 2060 4-bit 7B models
x86-cuda-8gb RTX 2070, RTX 3060 Ti, RTX 4060 Comfortable 7-8B models
x86-cuda-12gb RTX 3060, RTX 4070 7-14B models, fast inference
x86-cuda-16gb RTX 4070 Ti, RTX 5060 Ti 14B+ models
x86-cuda-24gb RTX 3090, RTX 4090 Large models, multiple concurrent
NVIDIA (Vulkan, no CUDA)
x86-vulkan-4gb GTX 750 Ti, GTX 950, GTX 960 Legacy cards, small models via Vulkan
x86-vulkan-8gb GTX 1070, GTX 1080 Medium models via Vulkan
AMD (ROCm)
x86-rocm-8gb RX 6600, RX 7600 Entry-level AMD GPU compute
x86-rocm-12gb RX 6700 XT Solid mid-range
x86-rocm-16gb RX 7800 XT Comfortable for 14B models
x86-rocm-24gb RX 7900 XTX Large models
CPU Only
cpu-only Any device Smallest quantized models, slowest

Hardware is auto-detected on first boot. The platform adapts to what's available — users with more RAM or better accelerators get access to larger models automatically.

Resource Overhead

Platform overhead (without models or agents): ~345 MB RAM

Component RAM
Armbian base ~200 MB
TinyAgentOS (FastAPI) ~67 MB
incusd (container management) ~67 MB
Metrics + health monitor ~5 MB

Investigating

  • Outlines — structured generation for reliable tool calling with small models
  • DSPy — automated prompt optimization for small local LLMs

Development

pip install -e ".[dev]"
pytest tests/ -v          # 123+ tests

Roadmap

  • Pre-built Armbian OS images for supported SBCs
  • Local assistant LLM (chat-based setup agent)
  • Cloud services (tinyagentos.com, agent email, subscriptions)
  • AI-aware desktop environment with auto-attaching MCP servers
  • Custom domain support for agents

Support the Project

TinyAgentOS makes AI agents accessible on affordable hardware.

  • Contact: jaylfc25@gmail.com
  • Donate: Buy Me a Coffee
  • Hardware donations/loans: We test on real hardware. If you have spare SBCs, GPUs, or dev boards and want to help expand compatibility, reach out at the email above.

License

MIT

About

Self-hosted AI agent platform for low-power hardware. App store, model management, agent deployment. Like Umbrel but for AI agents.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • Python 78.0%
  • HTML 18.9%
  • Shell 1.6%
  • CSS 1.5%
Morty Proxy This is a proxified and sanitized view of the page, visit original site.