After MCP, A2A, & AG-UI, there’s another agent protocol:

HarnessRouter is a new agent protocol that allows developers to build products using harnesses without worrying about harness engineering. This infrastructure layer turns existing harnesses like Codex and Claude Code into plug-and-play agent backends, enabling seamless task execution, result retrieval, and harness switching through a single API.

The Unified Harness Protocol (UHP) and OpenAI Responses-compatible API are provided, handling persistent sessions, streaming progress, files, artifacts, cancellation, and structured failures. Eight harness-model configurations can be run simultaneously on the same task, with each metric comparing its best and worst results. The lowest-cost and fastest configurations may vary depending on the task.

To get started, users can run HarnessRouter locally, integrate it into their product, use starter kits, or utilize a managed Cloud Self-host solution. The process begins with a single Docker command, followed by connecting a model provider and running the first task. No HarnessRouter account or bundled model or trial key is required. Docker pulls the image if necessary, and a named volume preserves data such as the database, files, installed harness CLIs, and workspaces between restarts.

To use HarnessRouter, users must have Docker and approximately 4 GB of disk space. A provider API key is needed, but no HarnessRouter account is necessary. After pulling the latest image, the upgrade and backup guide should be followed. If port 3000 is busy, users can use -p 127.0.0.1:3100:3000 and open port 3100 instead. The Console and Gateway run unprivileged, while agent processes run as their session's user.

Once the instance is running, users can access it by signing in at http://localhost:3000 or the chosen host port. They can change the default password in Profile and keep the instance local until it is changed. To use a custom harness, users must first add a provider integration, choose a supported harness, and select New task. They can then pick an available model and provide a concrete task for the agent to execute. Built-in harnesses can be used without this step.