Open-source infrastructure for agent-native research, AI safety evaluation, and trustworthy autonomous systems.
QitOS is a research-driven open-source organization building composable tools for the AI agent ecosystem. Our focus:
- Agent-native research infrastructure — frameworks designed around the agent lifecycle, not web scaffolding
- Safety evaluation and risk discovery — reproducible benchmarks, signal detection, and forensics for frontier AI
- Reusable open-source tooling — small composable kernels over monolithic platforms
- Research-friendly defaults — observability, replay, and reproducibility built in from the start
We build frameworks, not products. Everything is designed to be inspected, extended, and recomposed.
|
The torch-flavor framework for agent researchers. One Status: Core Framework · Agent orchestration · Benchmark harness · Trajectory inspection |
A safety evaluation framework for AI agents. Reproducible, observable, and retryable evaluations across benchmarks and environments. Status: Core Framework · Safety evaluation · Benchmark adapters · Risk analysis |
|
|
Applications and showcase agents built with QitOS. Demonstrates how the core framework powers real agent systems. Status: Zoo · Showcase agents · Application templates · QitOS integration examples |
An AIGC forensics agent for detecting AI-generated content. Multi-modal analysis with evidence-backed verdicts and reproducible reports. Status: Research Tooling · AIGC forensics · Evidence analysis · Multi-modal detection |
Official benchmark adapter collection for Snowl. 20+ safety and capability benchmarks as pluggable adapters. Status: Evaluation · Benchmark adapters · Safety benchmarks · Registry integration |
- Research-first, production-aware — Designed for researchers who need observability and reproducibility, not just demo scaffolding.
- Composable by design — Small kernels that compose. No framework lock-in, no hidden orchestration.
- Safety and evaluation built in — Benchmarks, signal detection, and forensics are first-class concerns, not afterthoughts.
- Minimal core, rich ecosystem — The core stays small. Applications and domain agents live outside it.
- Open benchmarks and reproducible experiments — Every run leaves a trace. Every evaluation can be replayed.
| Resource | Link |
|---|---|
| QitOS Framework | github.com/Qitor/qitos |
| QitOS Docs | qitor.mintlify.app |
| Safety Evaluation | github.com/Qitor/snowl |
| Agent Zoo | github.com/Qitor/qitos-zoo |