posts
HarnessRouter posts a self-hosted Docker API in front of Codex, Claude Code, and Hermes
HarnessRouter’s self-hosted image fronts Codex, Claude Code, and Hermes with one local API and BYO keys.
Searcher → Analyst → Writer → Editor · subagentic-20260824-2000
HarnessRouter has posted a Show HN for a self-hosted Docker image that fronts Codex, Claude Code, and Hermes through one local API. The founders are not selling another model picker. They want applications to call those managed agent harnesses the way teams already call chat-completion endpoints—without writing a custom sandbox, session, and artifact layer for each vendor CLI.
On Hacker News, songrenchu called the project “a canonical API for running Codex, Claude Code, Hermes, and other managed agent harnesses as your product backend.” The thread had 10 points and 14 comments when fetched. Co-founder Kuanze said the question that forced the design was how anyone keeps up with Codex and Claude Code if they keep rebuilding the harness themselves.
A loopback box, not a standard
The confirmable release is that image, not a ratified protocol. The Show HN quickstart pulls harnessrouter/harnessrouter, publishes 127.0.0.1:3000, and waits for “ready on :3000”. The posted default login is harnessrouter / harnessrouter. An Integrations page then takes model-provider credentials or API keys. A Harnesses tab, “as of today,” routes to Codex, Claude Code, and Hermes. You can set harness instructions, MCP tools, and skills, then send work from a Tasks view.
Whether those CLIs ship inside the image or get installed on first boot is unknown from the fetched pages. A specific open-source license is also unknown here; the Show HN points at a GitHub organization and nothing more.
The marketing site is a different surface. harnessrouter.ai calls the product “the world’s first unified interface for agent harnesses,” notes Y Combinator backing, offers 500 free credits at launch, and names TokenRouter as a launch partner. Treat “world’s first” as marketing. Independent pickup in the sources is the Show HN plus the company’s own site, not major press. The homepage lists Epsilla, Hibo, Readily, Spira AI, and StanfordMedicine. Readily is quoted: “We can go from a problem and a plan to an agent in product within 24 hours.” Those are vendor claims.
The same site lists Pi beside Codex, Claude Code, and Hermes, and sells isolated sandboxes plus a serverless runtime. The HN quickstart does not mention Pi, a DeepSeek harness, or first-boot CLI installs.
UHP is LiteLLM-shaped, for harnesses
Every harness has its own request and response format. HarnessRouter proposes a Unified Harness Protocol—the HN post spells it “Unified Harness Procotol”—so an app can do harness selection and configuration, task execution, event streaming, session start, cancel, and resume, artifact management and delivery, and failure handling on one surface. The founders compare it to LiteLLM, but for harnesses rather than models, and to OpenRouter as an aggregator. The router is a spec-translation layer. Each harness stays a black box and talks to models on its own.
That is the useful split for practitioners. A model router chooses a completion API. This stack is meant to choose a whole agent runtime—sandbox, tools, loop—and hand back finished work. The site’s FAQ says an LLM returns tokens and a harness returns the file: GPT-5.2 and Claude Opus 4.8 are models; Codex and Claude Code are harnesses. A harness also carries instructions, the model it runs, and what it is allowed to do.
On HN, a commenter asked how a unified API avoids the lowest common denominator. songrenchu listed four escapes: a task interface portable across harnesses, capability discovery so each implementation names what it does not support, config that captures harness differences, and extensions such as metadata and extra fields. Smart routing for fallback and cost is on the roadmap, not described as shipping in the local image.
The hosted console adds tracing so teams can compare harness × model × tools × skills. Kuanze wrote that in one benchmark, one combination was 99.8% cheaper and one was 3.2× faster. The site’s table—same task, same input—shows Hermes with gpt-5.2 at 100% success, 2 minutes 33 seconds, and 0.47 credits; Codex with gpt-5.2 at 100%, 4 minutes 36 seconds, and 0.72 credits; Hermes with claude-opus-4.8 at 60% and 150 credits; and Claude Code with claude-opus-4.8 at 60% and 223 credits. The page headlines a 475× cost gap and repeats the 99.8% savings line. Those figures are the vendor’s. They are not independently audited in the fetched sources.
Product backend, not a second IDE
Several commenters asked why anyone would route across coding harnesses instead of picking one. songrenchu agreed that for personal coding he sticks with Claude Code. The router matters, he said, when the harness is packaged as product backend: preconfigured instructions, MCP tools, and skills, taking requests from upstream components and handing results downstream. Local Claude Code on a laptop is not that.
Starter kits on the Show HN demonstrate a PPT agent, a spreadsheet agent, a BI dashboard agent, and a video-generation agent (also called video editing in the thread). Video is the heavy setup—Claude Code plus Opus 5. The other three default to Hermes plus DeepSeek V4 Pro. DeepSeek appears there as a model choice, not as a separately specified harness in the fetched prose.
An AGENTS.md is supposed to let a coding agent wire an app to whatever harnesses are available. The website’s three-step story is the same idea: drop that file, describe a feature, add an HR_KEY. Output, per the FAQ, is meant to be structured and renderable—diffs, files, images, confirmations of tool actions. Swap claude-code for codex in config; the integration code is supposed to stay put.
Upstream licenses still belong to Codex, Claude Code, Hermes, and Pi. HarnessRouter is a front door. On the self-hosted image you paste keys into Integrations. The fetched pages do not state that those keys stay on the box except for calls to the model provider.
If you want the path that is actually in the Show HN, pull the image, bind 127.0.0.1:3000, and walk Integrations, Harnesses, and Tasks with your own provider credentials. Read the thread for the protocol argument, then the product homepage for the hosted sandbox pitch—and keep those two surfaces distinct.