OpenRany

1. Abstract: What is an Agent OS?

Traditional agent frameworks are chatbot wrappers. They wait for the user to type a message, call an LLM, return a response, and wait again. They do not take initiative, do not schedule autonomous work, and do not maintain persistent state between invocations.

OpenRany is fundamentally different. It is an Agent Operating System — a persistent runtime that manages autonomous agents as first-class processes with their own lifecycles, schedules, budgets, and capability sets. Like an operating system manages CPU processes, OpenRany manages agent processes: scheduling them, isolating them, giving them tools, and enforcing resource limits.

OpenRany compiles to a single ~32 MB binary. One install. One command. Agents are live.

curl -fsSL https://openrany.orzatty.com/install | sh
openrany init     # Walks through LLM provider setup
openrany start    # Dashboard live at http://localhost:4200
openrany hand activate researcher  # Remote research agent starts working

2. 14-Crate Architecture

OpenRany is a Rust Cargo workspace of 14 crates, each with a single architectural responsibility. The modular structure enables independent testing, feature flagging, and compilation:

3. 7 Autonomous Hands

Hands are OpenRany's primary user-facing innovation — pre-built autonomous capability packages that run independently on schedules, without requiring user prompts. Each Hand is defined by a HAND.toml manifest specifying tools, settings, requirements, and dashboard metrics, plus a multi-phase operational playbook (500+ word expert procedures, not one-liners).

openrany hand activate researcher  # Starts working immediately
openrany hand status researcher    # Check progress
openrany hand activate lead        # Daily prospecting on schedule
openrany hand pause lead           # Pause without losing state
openrany hand list                 # See all available Hands

4. 16 Security Systems — Defense in Depth

OpenRany was designed security-first. Every security layer is independently testable and operates without a single point of failure. The 16 systems operate simultaneously and provide comprehensive coverage:

5. 40 Channel Adapters

OpenRany agents can deliver results and accept commands from 40 messaging platforms, across every major category:

Each adapter supports per-channel model overrides, DM/group policies, rate limiting, and output format customization (e.g., HTML for Telegram, Markdown for Discord).

6. 27 LLM Providers — 123+ Models

OpenRany routes through 3 native LLM drivers (Anthropic, Gemini, OpenAI-compatible) to reach 27 providers:

Anthropic, Gemini, OpenAI, Groq, DeepSeek, OpenRouter, Together, Mistral, Fireworks, Cohere, Perplexity, xAI, AI21, Cerebras, SambaNova, HuggingFace, Replicate, Ollama, vLLM, LM Studio, Qwen, MiniMax, Zhipu, Moonshot, Qianfan, Bedrock, and more.

Features: intelligent routing with task complexity scoring, automatic fallback on error, cost tracking per-request, and per-model pricing in the model catalog. Agents can specify model preferences per task type — e.g., use deepseek-r1 for reasoning tasks, gemini-flash for speed-critical tasks.

7. Performance Benchmarks

All data from official documentation and public repositories — February 2026. OpenRany positions as the optimal balance point between ZeroClaw (minimal, tiny) and Python-based frameworks (feature-rich but heavy):

Cold Start Time (lower is better)

Idle Memory (lower is better)

Security Systems (higher is better)

8. OpenAI-Compatible API

OpenRany exposes a drop-in replacement for the OpenAI Chat Completions API. Any tool built for OpenAI (LangChain, LlamaIndex, custom integrations) works with OpenRany by changing one URL:

// Drop-in replacement — change only the base URL
curl -X POST localhost:4200/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "researcher",     // This is an OpenRany Hand, not an LLM model
    "messages": [{"role": "user", "content": "Analyze Q4 AI agent market trends"}],
    "stream": true
  }'

140+ REST/WS/SSE endpoints covering: agents, memory, workflows, channels, models, skills, A2A (Agent-to-Agent), Hands, analytics, and administration.

9. Migration Engine

Switching from an existing agent framework to OpenRany requires one command. The migration engine imports agents, conversation history, skills, and configurations from OpenClaw, LangChain, and AutoGPT:

# Migrate everything from OpenClaw
openrany migrate --from openclaw

# Dry run first to preview changes
openrany migrate --from openclaw --dry-run

# OpenRany reads SKILL.md natively and is compatible
# with the ClawHub marketplace — no skill rewrites required

10. Roadmap