What is OryxOS

OryxOS is an open-source Agent Harness OS: an Agent runtime you install on your own Kubernetes cluster, VMs or bare metal. It gives every business Agent (ops, support, knowledge, sales assistants and more) shared model access, a reasoning loop, memory, tool calling, sandboxing and audit.

One directory defines an Agent; one runtime hosts them all. Self-hosted — your data never leaves.

Teams do two things: write an Agent directory (a single AGENT.md) and wire up a few tools. Where messages come from, how the LLM is called, how context carries over and how audits are recorded — that is OryxOS's job.

Why OryxOS

Every company has work that belongs to Agents, yet most Agents never leave the demo. They get stuck on four barriers:

Barrier	Today	The OryxOS way
Defining an Agent takes code	The people who know the business can't do it	One AGENT.md, written in natural language
Cloud platforms take your data	Regulated industries can't pass compliance	Self-hosted; data stays on your infrastructure
Execution is a black box	No audit, no allow-lists — nobody ships it	Mandatory sandbox; every LLM and tool call is audited in the database
Running a fleet is hard	Nobody provides an OS layer for many Agents	One runtime manages the lifecycle and shared services of many Agents

The deeper point: what keeps Agents from working reliably in production is rarely the model — it is the environment they run in. OryxOS is not yet another Agent; it is the runtime that lets a fleet of Agents run reliably.

Agent runtime vs. Agent Harness OS

• Agent runtime: the execution kernel that runs a single Agent — calling the model, executing tools, managing context, controlling the reasoning loop.
• Agent Harness OS: sits above the runtime and manages a fleet — Agent lifecycles, shared inbound channels and outbound integrations, shared memory, multi-tenancy and governance.

In operating-system terms, the runtime is a process's execution environment; the Agent Harness OS is the layer that manages many processes, schedules resources and provides shared services. A runtime runs one Agent; OryxOS runs and manages all of them.

The north-star formula:

Natural language (AGENT.md) + Memory + Tools + MCP + Skills + Knowledge + Notify = an Agent

Five core capabilities

Capability	Description
LLM access	A provider abstraction over DeepSeek, Qwen, Kimi and other mainstream models; Agents never see the vendor. Multiple providers are distinguished by explicit mapping; local inference is supported
ReAct loop	A home-grown reasoning engine: the LLM decides whether to call a tool → OryxOS executes it and feeds back the result → reasoning continues until an answer or the iteration limit
Memory	Session memory plus long-term memory; long-term memory lives in a human-readable MEMORY.md with keyword search, with room to upgrade to vector search
Tools	Nine built-in tools (files, shell, HTTP, memory, notify); three extension tiers: zero-code Agent directory + MCP, light-code custom MCP server, full-code @Tool Bean
Web Service	Every capability is exposed over a REST API, so any language can integrate over HTTP

Key features

• 🤖 One directory = one Agent: any directory with an AGENT.md is an Agent — no code, many Agents per instance
• ☕ Java-native: JDK 21 + Spring Boot 3, a single executable JAR that fits your existing Java toolchain
• 🔒 Under your control: runs on your infrastructure, data never leaves, no cloud lock-in
• 🛡️ Secure by design: file, command and network allow-lists on every tool call; secrets via environment variables; end-to-end audit
• 🧠 Home-grown ReAct: the core loop is implemented in-house, not borrowed from an Agent framework
• 🔌 Open standards: MCP for tools, A2A for collaboration, Agent directories modeled on Anthropic Agent Skills
• ⏰ Scheduled runs: Agents can run on a cron schedule and push results to your team IM
• 🌐 Stateless and scalable: stateless instances with externalized state, ready for a distributed future

Design principles

• The runtime over any single Agent: the key deliverable is an environment where any Agent runs reliably
• Own the core: the reasoning loop is home-grown; model protocol adapters reuse mature libraries
• Configuration is the Agent: an Agent is defined by configuration, not code
• Open standards: interoperate with the ecosystem instead of inventing protocols
• Stateless instances, external state: the prerequisite for scaling out
• Security is the foundation, not a patch: controlled sources, least privilege, mandatory sandbox, no plaintext secrets, full audit
• Deliberate phasing: get the single-node kernel right first; every architectural upgrade must be justified by real usage

How OryxOS relates to other projects

Category	Examples	Relationship
Agent OS	OpenClaw, Hermes Agent	Same category, different focus: they target individuals and small teams; OryxOS targets regulated enterprises
Orchestration platforms	Dify, Coze	Complementary: they can sit on top and call OryxOS's API
Agent frameworks	Spring AI, LangChain4j	Reused: OryxOS's LLM layer is built on Spring AI Alibaba

In short: a framework hands you materials to build a house; an orchestration platform scripts a workflow; OryxOS hands you a move-in-ready, governable, auditable home for your Agents.

Next steps

• Quick Start: build from source and see how 1.0 will be used
• Architecture: layers, key decisions and modules
• Roadmap: from a single-node kernel to distributed Agent collaboration