HarnessRouter

HarnessRouter lets teams put complete agent harnesses behind their own product through one API. Instead of building sandboxes, sessions and recovery yourself, you pick a harness and model, add skills and instructions, and send tasks. It's built for teams turning an agent demo into something users can rely on, and it works as a managed alternative to Claude Managed Agents.

It's not a model router. It runs the whole harness: the execution loop, tools, files and working environment.

• Multiple harnesses: Codex, Claude Code, Hermes Agent, Pi and DeepSeek Harness, switchable without changing your product API.
• Unified Harness Protocol: an open, versioned contract for running tasks, streaming progress, continuing sessions, exchanging files and cancelling work. Its task surface is compatible with the OpenAI Responses API.
• Serverless runtime: one isolated sandbox per task, with session handling, recovery, artifacts and execution traces.
• Benchmarking: compare harness and model pairs on your own tasks for quality, latency and cost. In one measured workload, changing only the harness cut cost by 1.5x to 2.1x.
• Migration path: move an existing agent backend with staged verification and rollback.

You keep your own interface, authentication, permissions and data.

HarnessRouter Cloud is the managed service. It starts free, and you're billed only while the agent actively works. Waiting costs nothing. You can use a HarnessRouter key or bring your own provider keys. Community Edition is the Apache 2.0 self-hosted version that runs the same protocol on your infrastructure, with your keys, state and files staying with you.