HarnessRouter

HarnessRouter sits between your product and the coding-agent harnesses that actually do the work. Instead of integrating Codex, Claude Code and Hermes separately — each with its own session model, streaming format, file handling and failure modes — you call one API and pick the harness with a configuration value.

Why a router

A harness is not a model. It is the loop around the model: the sandbox, the tool calls, the retries, the artefacts it leaves behind. Products that embed agents end up rebuilding that loop per vendor, and then rebuilding it again when a better harness ships. HarnessRouter's answer is the Unified Harness Protocol, a single request/response shape that covers session lifecycle, streaming, file upload, workspace persistence, cancellation and failure.

Supported harnesses

Codex (code, apps, images), Claude Code (code, files, documents) and Hermes, an open-source autonomous agent that runs against any frontier model. A fourth, Pi, is listed as coming soon.

Two ways to run it

• Community Edition — Apache 2.0, ships as a single Docker container with local SQLite state and a management console. Self-hosted, free, and the reference implementation of the protocol. Source at github.com/HarnessRouter/harnessrouter.
• HarnessRouter Cloud — managed serverless execution that provisions isolated sandboxes on demand and handles retries, timeouts, permissions and cost caps. A free tier includes 500 launch credits, with Developer, Production and Scale plans above it.

Benchmarking

The platform ships built-in benchmarking so you can compare harness/model pairs on the same task by success rate, latency and cost. The vendor cites a case where the cheapest passing configuration cost 99.8% less than the most expensive one on identical work — the kind of spread that makes a router worth having independently of vendor lock-in.

Who uses it

Y Combinator-backed, with Epsilla, Hibo, Readily, Spira AI and Stanford Medicine named as users on the site.