01Provider
Your choice of models
Connect model providers through explicit mappings, including your own internal inference services.
Spring AI · Spring AI Alibaba
Java-native · Self-hosted · Agent Harness OS
OryxOS
One foundation. A world of Agents.
Define an Agent with a directory. Models, reasoning, memory, tools, and APIs share one Java-native runtime. Describe the work in Markdown. Let OryxOS run it.
JDK 21 · Spring Boot · ReAct · MCP · REST API
In development preparation · Target architecture and capabilities
WHY ORYXOSBeyond a conversation.
Beyond a conversation.
A foundation for action.
A business Agent needs more than a model call. It needs context, tools, memory, and a reliable way to run.
OryxOS brings those shared capabilities into one runtime so every Agent can reuse them and teams can focus on their work.
Explore the design principlesYOU DEFINE01
- The task and the Agent’s instructions
- Models, tools, and shared Skills
- Business integrations and trigger rules
THE RUNTIME HANDLES02
- ReAct execution, context, and memory
- Tool execution, allowlists, and audit records
- Shared CLI, API, and scheduled entry points
One engine. All the capabilities.
Three ways to trigger. One execution path. A controlled core with clear interfaces.
Single-instance target design · Spring AI adapts protocols. OryxOS controls reasoning and tool execution.
Five capabilities. One runtime.
Build a solid single-node runtime, then grow toward governance and distribution.
02ReAct
Own the reasoning loop
A self-implemented reasoning loop controls tool execution, message history, and iteration limits.
Reason → Act → Observe
03Memory
Keep the context
Access session history and long-term memory through one service. Keep preferences and facts in MEMORY.md.
Session · MEMORY.md · Recall
04Tool
Connect to real systems
Built-in file, HTTP, shell, and notification tools, extended with MCP servers or native Java tools.
Built-in · MCP · Java @Tool
05Web Service
Bring Agents to your app
A REST interface for sessions, Agent invocation, memory, and tool information — usable from any language.
REST API · CLI · Scheduler
CONFIGURATION IS THE AGENTA Markdown file.
Explore Agent and Skill design →
A Markdown file.
A business Agent.
A directory defines the Agent: frontmatter describes how it runs; the body describes what it does. Bind shared Skills and load their content on demand.
01Describe the task
02Choose models and tools
03Set the trigger
Design examples, not instructions for a released runtime.
markdownDESIGN EXAMPLE
---
name: daily-tech-digest
description: 每日科技日报
provider:
name: deepseek
model: deepseek-chat
tools: [http_get, read_file, notify]
schedules:
- cron: "0 0 9 * * *"
timezone: Asia/Shanghai
message: "生成今天的科技日报"
---
关注 AI 与芯片领域，汇总今天的重要新闻。
按已绑定的 digest-format Skill 整理内容，
标注信息来源，通过 team-lark 渠道发送。Start with everyday work.
Validate the foundation with practical tasks that bring models, memory, and tools together.
Daily weather briefing
Fetch the weather on schedule and send an outfit recommendation to your team. Manual runs use the same path.
Core acceptance scenarioA digest that remembers
Use remembered interests, fetch news through MCP, and format the digest with a bound Skill.
Core acceptance scenarioGitHub project digest
A dedicated tool wraps the data-gathering script; the model summarizes the results.
Additional design scenarioEngineering & operations
Connect engineering, monitoring, and ticketing tools to support multi-step operational tasks.
Application directionEnterprise knowledge
Combine business knowledge with tools to support research, documentation, and reusable team expertise.
Application directionAI inside your workflow
Invoke Agents over HTTP with a persistent conversation or a single independent task.
Application directionThree ways to extend. Your choice.
{ }No-code
Markdown + existing MCP
Describe the task in AGENT.md, bind shared Skills, and reuse existing MCP servers.
AGENT.mdSKILL.mdMCP
↔Low-code
Your own MCP server
Wrap business APIs in your preferred language to connect ERP, CRM, monitoring, or developer tools.
PythonTypeScriptJavaGo
</>Native integration
Native Java tools
Expose Spring Bean methods through @Tool and reuse your existing Java services and dependencies.
Spring Bean@ToolOryxTool
An HTTP interface for your business.
Ten REST endpoints planned for the core phase, under /api/v1.
POST
/sessionsCreate a sessionPOST
/sessions/{id}/messagesSend a messageGET
/sessions/{id}Read conversation historyDELETE
/sessions/{id}Archive a sessionPOST
/agents/{name}/invokeInvoke an AgentGET
/profilesList Agent profilesGET
/memoryRead long-term memoryGET
/toolsList available toolsGET
/healthHealth checkGET
/infoRuntime informationStart focused. Grow deliberately.
Deliver a usable core, then let real needs drive the architecture.
Current focus
01Single-node runtime
Five core capabilities, multiple Agents, CLI / REST / scheduled triggers, persistence, and basic audit records.
Planned
02Distributed foundation
Externalized state, stateless instances, and replicas for high availability and horizontal scaling.
Vision
03Cross-node collaboration
An Agent communication layer with A2A integration for discovery, delegation, and reliable async collaboration.
Channels, knowledge, observability, web management, and governance evolve with actual needs. The roadmap is not a release commitment.
BUILD IN THE OPEN
Build the Agent foundation with us.
Start with a problem you want to solve. Read the design, discuss an idea, or make your first contribution.
terminal
$ git clone https://github.com/lt950303/agent-lean-demo.git
$ cd agent-lean-demoCurrently in design and development preparation. Runtime packages are not available yet.