Build enterprise AI applications. Govern every AI action.
An open-source, self-hosted Agent Harness OS for the enterprise.
OryxOS is an open-source, self-hosted Agent Harness OS for enterprise AI applications. It lets enterprises run any Agent inside controlled context, execution environments, Tools, permissions, acceptance criteria, and governance. Build R&D automation, knowledge support, ticket collaboration, operations diagnostics, and other domain applications on one shared runtime.
OryxOS is the operating foundation for enterprise AI applications. Its direction is to give enterprises one place to build, run, govern, and optimize their own AI applications. Application teams compose Agents, models, Knowledge, Tools, MCP servers, and Flows to solve business problems; OryxOS supplies the execution semantics that make those applications consistent, extensible, and operable across the organization.
Agent Harness OS describes the shared Runtime and governance semantics around every Agent and Task. Agents are replaceable executors; the enterprise Harness — its context, environment, capabilities, policies, evidence, and operational feedback — is the asset that compounds.
Its core principle is simple: applications can freely compose capabilities, but they cannot bypass governance. Every Task runs through a common path for identity, authorization, policy, approval, model routing, Tool execution, audit, trace, cost attribution, and state updates.
Task → Context / Environment → Identity / Authorization → Policy / Approval
→ Model / Tool Execution → Artifact / Evidence → Audit / Trace / Cost → State Update
OryxOS is the shared foundation beneath enterprise AI applications: a stable kernel with extensible capabilities and a governance path that remains part of execution.
| Capability | What it provides |
|---|---|
| Agent and Task Runtime | A consistent lifecycle for Agent execution, context, controlled environments, sessions, memory, state, artifacts, evidence, and durable work. |
| Model routing and cost optimization | Selects models according to task requirements such as quality, latency, availability, cost, and data boundaries; evolves toward fallback, budgets, quotas, and organization-wide optimization. |
| Knowledge and context | Knowledge bases, retrieval, long-term memory, and extensible context assembly for business applications. |
| Tools, MCP, and enterprise integrations | Reusable capabilities for acting on enterprise systems through built-in Tools, MCP servers, Skills, channels, and custom extensions. |
| Governance Runtime | Identity, authorization, policy, approval, sandbox boundaries, and audit are part of the action path rather than after-the-fact reporting. |
| Operations and reliability | Traceability, acceptance evidence, cost attribution, evaluation, recovery, retries, observability, and the foundation for dependable production operation. |
- Build business applications, not infrastructure. Teams focus on their domain while reusing a common AI runtime, execution environment, and capability layer.
- Govern every AI action. Models and Tools can be composed by applications, but policy, approval, audit, and security boundaries travel with execution.
- Optimize as a system. Routing, cost, evidence, trace, evaluation, and operational data become shared enterprise assets rather than isolated application concerns.
Available today — single-node runtime kernel
Define and run Agents on your infrastructure with model Providers, Memory, Knowledge, built-in Tools, MCP, Skills, notifications, scheduling, sandbox controls, audit records, REST APIs, and a Web Manager.
Evolving — enterprise execution foundation
- Durable Tasks — stable lifecycle, checkpoints, artifacts, waiting, cancellation, retry, and recovery.
- Model routing and cost optimization — policy-aware routing, fallback, budgets, quotas, usage attribution, and data-boundary controls.
- Governance Control Plane — identity context, policy distribution, approval workflows, organization boundaries, and administrative lifecycle.
- Operational intelligence — cost analysis, trace analysis, evaluation, regression feedback, and cross-application optimization.
- Distributed foundation and Flows — shared storage, stateless instances, high availability, and governed workflow execution.
See the Vision & Roadmap for the evolving technical plan.
git clone https://github.com/oryx-labs/oryxos.git
cd oryxos
mvn package -DskipTests
bin/start.shThe default setup starts with the built-in mock Provider and does not require an LLM API key. Open http://localhost:8080/admin/ after startup.
For installation, deployment, provider configuration, API references, and development instructions, see docs.
- License: Apache License 2.0
- Contribution: CONTRIBUTING.md
- Long-term aspiration: grow an open community and pursue incubation at the Apache Software Foundation.