HarnessRouter/harnessrouter

What it solves

HarnessRouter eliminates the need for developers to build custom integration logic for every different agent harness (such as Codex or Claude Code) they want to use in their product. It provides a single, unified API layer that allows products to run tasks, manage sessions, and switch between different agent backends without rewriting their integration code.

How it works

It implements the Unified Harness Protocol (UHP) and offers an OpenAI Responses-compatible API. The system consists of a Console for management, a Gateway for API handling and harness lifecycle, and a Runner that executes harnesses within isolated session workspaces. Users can connect their own model provider keys, configure built-in or custom harnesses (including instructions, tools, and MCP servers), and trigger agent tasks via API calls using a harness_id to specify the backend.

Who it’s for

Developers and product teams building agentic applications who want to use professional agent harnesses as pluggable backends while maintaining control over infrastructure, costs, and latency.

Highlights

• Unified Interface: One API to manage multiple agent harnesses, reducing integration overhead.
• OpenAI Compatible: Works with existing OpenAI Responses SDKs and streaming parsers.
• Session Management: Handles persistent sessions, streaming progress, file attachments, and task cancellation.
• Customizable Harnesses: Support for creating custom harnesses with specific instructions, tools, and skills.
• Flexible Deployment: Available as a self-hosted Community Edition via Docker or as a managed Cloud service.