HarnessRouter

About HarnessRouter

HarnessRouter is an API and runtime for running AI agent harnesses as a product backend. It provides a unified interface, called the Unified Harness Protocol, that lets a developer send a task and receive streamed progress, results, and artifacts regardless of which underlying harness and model executes it. The platform runs each task in its own isolated sandbox and supports swapping between harnesses and models through configuration rather than separate integrations. It lists 30 supported harnesses, including Codex, Claude Code, Hermes, Pi, DeepSeek Harness, Kilo Code, and Qwen Code.

The product includes a benchmarking tool (Harness Arena) for comparing harness and model combinations on cost, speed, and accuracy for a given task, along with published harness rankings. Included features cover one sandbox per task, session handling, full execution traces, and defined failure/terminal states rather than silent retries or failovers.

It is aimed at teams building agent-powered products who want to avoid building and operating their own harness infrastructure, including teams without a dedicated infrastructure team. Access is provided via an API key, with a console for configuring harnesses and a migration guide for moving an existing agent's execution onto HarnessRouter. It has an open-source Community Edition and is backed by Y Combinator.

Key features

• Unified API to run and switch between 30+ agent harnesses (e.g. Codex, Claude Code, Hermes, Pi)
• Isolated sandbox per task with full execution trace and artifacts
• Smart routing selects a compatible harness/model configuration by quality, cost, or latency
• Harness Arena for comparing time, cost, and accuracy across harness/model pairs on the same task
• Serverless provisioning of sandboxes that scales automatically without customer-operated infrastructure
• Defined failure contract: every run ends in a clear terminal state (e.g. timeout → incomplete) with no silent auto-retry

Frequently asked questions