HarnessRouter

HarnessRouter puts existing agent harnesses such as Codex, Claude Code, and Hermes behind one API, so a product can run tasks, retrieve results, and switch harnesses without integrating each one separately. The API is compatible with the OpenAI Responses API, and the harness is chosen per request through a harness id in the request metadata.

It implements the Unified Harness Protocol (UHP), a public, versioned contract that defines harness selection, persistent sessions, streaming progress, files, cancellation, structured failures, and harness-managed tools and skills. The repository holds the reference implementation of UHP, its OpenAPI and JSON Schema definitions, and a conformance suite.

Community Edition is self-hosted as one Docker deployment with three parts: a Console for the UI and API entry point, a Gateway that serves the Responses API and manages the harness lifecycle, and a Runner that runs harnesses in separate session workspaces with their own operating-system users. You connect your own model provider keys, and provider keys, sessions, files, and workspaces stay on your infrastructure. A managed HarnessRouter Cloud runs each task in its own serverless, isolated sandbox through the same API contract, and a custom harness configured locally can be uploaded to it.

Custom harnesses start from a built-in base harness and add agent instructions, tools (including MCP servers), skills, and a default model. Starter kits show harnesses as backends for slides, sheets, dashboards, videos, and a multi-agent workplace. A benchmark of eight harness and model configurations on one task reports cost differing from 223 to 0.47 credits and latency from 4m 36s to 1m 25s between the best and worst runs.