HarnessRouter on Railway
One Responses-compatible API and web console for agent harnesses like Codex and Claude Code, on Railway.
HarnessRouter Community Edition puts agent harnesses such as Codex, Claude Code, Hermes and Pi behind one OpenAI Responses-compatible API and a web console. You call one endpoint and choose the harness per request.
It is aimed at developers adding agent features to a product, teams comparing harnesses on the same task and files, and people building and testing custom harnesses with instructions, skills, MCP tools and environments.
The Community Edition is released under Apache-2.0. The agent CLIs it installs each keep their own terms.
What HarnessRouter can do
- OpenAI Responses-compatible API with harness selection per request
- Web console with API key management
- Runs each agent CLI as its own Linux user in its own workspace
- Background and streaming responses
- Bring-your-own provider keys, encrypted at rest
- Starter kits for slides, sheets, dashboards and videos
What people use it for
- Adding agent features to a product through one API
- Comparing harnesses on the same task and model
- Building and testing custom harnesses
- Running starter kits on your own infrastructure
Videos
Running HarnessRouter on Railway
The Railway template deploys a single HarnessRouter service from the upstream image, pinned by digest, with a volume at /data for its SQLite databases, files and the agent CLIs installed on first start. A console login password is generated at deploy, and you add at least one model provider key. The first start takes a few minutes while the CLIs install. Cost depends on CPU and memory, which the template documents at roughly 0.5 to 1 GB of RAM per running turn, plus the volume.
HarnessRouter templates on Railway
Tracklayer templates are listed first. Other HarnessRouter templates follow, ordered by total deployments on Railway. How this works.
Deployment numbers as of 2026-11-20.
Questions about HarnessRouter
How do I sign in?
Open the public URL and sign in as harnessrouter with the HR_AUTH_PASSWORD value from the service's Variables tab. You can set your own password under Profile.
Which model keys does it need?
At least one provider API key, such as Anthropic, OpenAI, Google AI Studio or OpenRouter. It works with API keys, not Claude or ChatGPT subscription logins.
Why does a long agent request time out?
Railway closes a request after 5 minutes without data and 15 minutes in total. Use background mode and poll for the result, or stream for shorter turns.
How much disk does it need?
The default set of agent CLIs is estimated at 1.5 to 2.5 GB and fits a 5 GB volume with room for workspaces. Only CLIs listed in HR_BACKENDS are installed.
Can I run several replicas?
No. The volume holds the databases and workspaces, so keep one replica and raise CPU and RAM instead.
Guides
Similar apps
Tracklayer maintains the templates marked as such; HarnessRouter is a trademark of its owners and this site is not affiliated with them.