Hey everyone 👋 Three weeks ago, we launched HarnessRouter here on YC with one promise:

Bring the world’s best AI agents into your app, with one API.

Today, we’re opening the layer underneath that promise.

We’re releasing HarnessRouter Community Edition and the Unified Harness Protocol (UHP). You can now run Codex, Claude Code, and Hermes as your product’s backend on infrastructure you control, using the same API contract as HarnessRouter Cloud.

TL;DR：

HarnessRouter is the world’s first unified interface for agent harnesses.

Your product sends a task through one agent API. HarnessRouter runs the supported harness you choose and returns completed work, including files, artifacts, progress, and results.

Sessions, streaming, cancellation, and failure handling follow the same contract across Codex, Claude Code, and Hermes. You can add or switch a supported harness without rebuilding your product integration.

🎥 Watch the launch video

Why agent harnesses aren't infrastructure yet

An agent harness is the runtime layer around a model. It executes tasks, uses tools, manages sessions and files, streams progress, and returns completed work.

Historically, product teams built this layer themselves: task execution, session management, streaming, file and artifact handling, and error recovery, then maintained everything as models and tools evolved.

Now frontier labs and open-source communities are building increasingly capable harnesses. Codex, Claude Code, Hermes, and DeepSeek Harness are examples. But they're still seen as coding tools rather than reusable infrastructure, so teams keep rebuilding this layer in-house. Just as product teams don't build their own databases, they shouldn't have to build their own harness layer either.

Even teams that do adopt an existing harness hit the next obstacle: every harness works differently, so each new product-harness pair needs its own integration. We call this the N×M harness problem. In practice, it becomes harness lock-in: teams stick with the first harness they integrated, even when better or cheaper options exist.

We asked a simple question: what if agent harnesses became plug-and-play infrastructure that any product can build on?

Plug-and-play means you integrate once, run any supported harness, and switch anytime a better or cheaper one ships.

That is what HarnessRouter makes possible.

What we’re open-sourcing

We’re releasing two core components under the Apache License 2.0:

• Unified Harness Protocol (UHP): UHP is an open standard that gives different agent harnesses a shared execution contract. This release includes a versioned specification, schemas, and a conformance suite covering harness configuration, task execution, progress, sessions, files, cancellation, errors, and results.
• HarnessRouter Community Edition: The self-hosted implementation of UHP. Gateway, Runner, and Console ship in a single Docker container. It runs Codex, Claude Code, Hermes, and more through one API and passes the full UHP conformance suite.

Your model-provider credentials, application state, and files remain under your control. Model calls still go to the providers you configure.

MCP connects agents to tools. UHP covers a different boundary: how an application drives a complete agent harness and receives its full execution lifecycle.

Start self-hosted. Scale with Cloud when you’re ready.

HarnessRouter Community Edition and HarnessRouter Cloud implement the same UHP contract, so one integration works for both. Start self-hosted on infrastructure you control, or scale into Cloud when you're ready.

HarnessRouter Cloud provides agent harness as a service (HaaS): fully managed harnesses with serverless execution in isolated sandboxes, so teams can scale concurrent runs without operating sandbox capacity themselves.

Agent harnesses can do more than write code: runnable examples

We also released a Starter Kit with runnable applications that show an end-to-end Community Edition integration.

One example puts a harness behind an Excel column so each row can trigger agent work. Other examples cover presentations, dashboards, and video generation.

Agent harnesses can do much more than write code. We want the Starter Kit to show what becomes possible when you put them inside products.

Starter Kit components are available under their applicable licenses.

Choose harnesses in production with evidence

We benchmarked eight Harness × Model configurations on the same task and input. Cost varied by 475× and latency by 3.2×. Even with the model fixed, changing the harness moved cost by 1.5–2.1× and latency by up to 1.95×.

Harness selection is a task-level production decision, not a one-time choice. HarnessRouter helps teams compare configurations on their own workloads. This was one controlled test, not a universal ranking; we published the methodology, results, and limitations.

What teams are already building

Teams are using HarnessRouter in very different settings:

• 🏥 Stanford Medicine is connecting its operational data through a shared agent layer, an “academic brain” for the organization.
• ✅ Readily.ai encodes domain expertise into agent skills and can move a new agent from problem and plan to production within 24 hours.
• 🎬 Spira.ai drives its generation pipeline through HarnessRouter.
• 🪄 hibo.ai uses HarnessRouter to execute and evaluate agent skills.

Why we built it

We’re Kuanze Ma and Richard Song, the co-founders of HarnessRouter.

Kuanze is a four-time founder focused on human-AI systems. Richard has spent more than a decade building large-scale systems, including at TigerGraph and Meta.

We both worked on agent systems before HarnessRouter. We kept seeing product teams rebuild the same application-to-harness layer, so we decided to turn that layer into shared infrastructure.

What we want from you

• If you build an agent harness or already maintain your own runtime, read UHP and tell us where it is wrong. Which parts should be standardized, and which should remain harness-specific? That comparison is worth more to us than a signup.
• Try the Community Edition Quickstart on a clean machine. Tell us where it breaks or where the execution lifecycle is incomplete.
• If your compliance requirements prevent you from using a hosted agent runtime, try Community Edition on infrastructure you control.
• If you want to contribute, we use a proposal-before-code process. Describe the problem, your proposed change, and its expected impact. You do not need to arrive with a finished implementation.
• Intros are welcome to teams shipping agent-powered features that do not want to own another backend.

Run it yourself

• Open-source repository
• Unified Harness Protocol
• Starter Kit
• Benchmark and methodology
• Docs and HarnessRouter Cloud

Reach us at contact@harnessrouter.ai.