On August 14, 2026, HarnessRouter, the world’s first unified interface for agent harnesses, open-sourced the Unified Harness Protocol (UHP) and HarnessRouter Community Edition. UHP is an open standard for connecting products to different agent harnesses; Community Edition is the protocol’s self-hosted implementation, an agent backend that runs out of the box.
An agent harness is the runtime layer around a model. Codex, Claude Code, and Hermes are harnesses: each lets the model use tools and skills, maintain sessions, handle files, stream progress, recover from failures, and deliver completed work.
Complete agent harnesses are becoming a distinct infrastructure layer. Labs such as OpenAI, Anthropic, and Nous Research, along with open-source contributors, continue to invest in these harnesses; capabilities that product and engineering teams once built themselves now come built in. Running a harness from a product, however, takes an AI agent API that covers the full execution lifecycle, not just model inference, and each harness answers this in its own way. Nothing transferred, so teams faced a choice: lock into one harness, or rebuild this runtime for every product. Until now, no standard has unified this interface.
HarnessRouter initiated that standard and provides the unified interface. Through one API, teams use complete agent harnesses as shared infrastructure, retain control over product logic and user experience, and can add or switch harnesses without rebuilding their backend.
“After experimenting with many different agents, we found HarnessRouter worked best for our situation. Work that could have taken weeks can now be done in a day, sometimes in hours. Beyond productivity, we are also seeing improvements in development quality,” said Jack Zeng, IT director at Stanford University School of Medicine.
The HarnessRouter Open-Source Release
This release makes that unified interface an open standard. The core protocol UHP and HarnessRouter Community Edition are released under the Apache 2.0 license.
Unified Harness Protocol (UHP)
UHP is an open standard, initiated and led by HarnessRouter, that defines a common way for products to work with different agent harnesses, covering harness selection and configuration, task execution, progress updates, session and file management, cancellation, error handling, and results.
The published UHP materials include a versioned specification and a conformance suite that checks implementations against it. UHP evolves through a public, versioned process that keeps the specification, the reference implementation (Community Edition), and the conformance suite aligned.
HarnessRouter Community Edition
HarnessRouter Community Edition is the protocol’s self-hosted implementation: an agent backend that runs out of the box on infrastructure developers control. It ships as a single Docker container packaging HarnessRouter Gateway, Runner, and Console; once deployed, locally or in production, a product can run Codex, Claude Code, and Hermes through one API. Model-provider credentials, application state, and files stay under the developer’s control, and the same interface is designed to support additional agent harnesses over time. Community Edition passes the full UHP conformance suite.
HarnessRouter also released an accompanying starter kit: runnable example applications that show end to end how a product integrates with Community Edition, which developers can extend into their own products under the applicable licenses.
The HarnessRouter Open-Source Community and How to Participate
Contributions are welcome. Substantial changes follow a proposal-before-code process: contributors first describe the problem, the proposed change, and its expected impact; once maintainers agree on the direction, the contributor or a maintainer implements it. Proposals and bug reports go through GitHub Issues, discussion happens in the HarnessRouter Discord community, and security vulnerabilities are reported privately.
HarnessRouter Cloud
Beyond the open-source release, HarnessRouter Cloud delivers agent harness as a service (HaaS): scalable serverless execution in isolated sandboxes. It manages concurrent runs, preserves sessions and files across runs, and streams progress. Teams can scale from a single run to multiple parallel runs without provisioning or maintaining sandbox capacity.
HarnessRouter Cloud records agent execution traces. Teams use them to configure, test, and compare combinations of harnesses, models, skills, and tools, and route each task type to the configuration that performs best on production cost, output quality, and latency. In a published HarnessRouter benchmark, eight harness-and-model combinations each ran the same task five times on identical input; cost per task varied by approximately 475×, with the lowest-cost successful run 99.8% below the most expensive; p95 end-to-end latency varied by more than 3×. Results vary by task.
Readily (YC S23), a healthcare compliance AI company, builds and ships the AI agents behind its product on HarnessRouter. “It feels like a plug-and-play solution in which we can take it and we plug it into our own use cases,” said Edward Tiong, co-founder and CEO of Readily. “HarnessRouter lets us run the best agent harnesses out of the box, so we can benefit from all the development the big labs are doing. We literally can go from a problem and a plan into an agent that is in prod within 24 hours.”
HarnessRouter Community Edition and HarnessRouter Cloud both implement UHP and expose the same API contract, so teams do not need separate product integrations for the two deployment models.
Founders and Company Background
“Product and engineering teams do not rebuild a database or train their own model for every application. They should not have to rebuild an agent harness for every agent-powered feature,” said Kuanze Ma, co-founder of HarnessRouter. “Adopt a complete agent harness like any other core infrastructure. Teams that stop rebuilding this layer are shipping product and winning customers; teams that don’t are still building infrastructure.”
While building hibo, a platform for generating and evaluating expert skills, Kuanze Ma designed and implemented an agent harness for reusable agent skills, and saw firsthand the heavy lifting of building and maintaining one. At HarnessRouter, he leads category and product strategy, developer adoption, and community building.
“Open source is how infrastructure earns trust: developers can read it, run it, and extend it,” said Renchu (Richard) Song, co-founder of HarnessRouter. “UHP is the specification anyone can read and contribute to. Community Edition is the software anyone can run and change.”
Renchu Song leads HarnessRouter’s unified execution architecture and production infrastructure. He previously worked on large-scale systems at Meta and TigerGraph. At Epsilla (YC S23), which he co-founded, he delivered agent harnesses as reusable infrastructure for product and engineering teams.
Developer and Community Resources
- Open-source repository: github.com/HarnessRouter/harnessrouter
- UHP specification: unifiedharnessprotocol.org
- Starter Kit: github.com/HarnessRouter/starter-kit
- Docs and cloud: harnessrouter.ai
- Discord community: discord.gg/nPcbwqVPb2
About HarnessRouter
HarnessRouter is the world’s first unified interface for agent harnesses. Through one API, products hand tasks to complete agent harnesses such as Codex, Claude Code, and Hermes and get completed work back, using them as the agent backend without rebuilding an agent runtime per product. HarnessRouter is available as the open-source, self-hosted Community Edition and as the fully managed HarnessRouter Cloud; both implement the Unified Harness Protocol (UHP), the open standard HarnessRouter initiated and maintains.