OpenRouter launches `create-headless-agent` for Bun-based multi-model CLIs

• OpenRouter shipped a new create-headless-agent skill that, according to OpenRouter's announcement, walks developers through building a customizable, multi-model headless CLI with Bun.
• The linked announcement post, as referenced in OpenRouter's post, frames create-headless-agent alongside OpenRouter's Agent SDK rather than as a standalone template.
• In OpenRouter's docs screenshot, the guide is positioned for CLI tools, API servers, and automation pipelines, with @openrouter/agent handling the inner loop.
• OpenRouter's docs link also points readers to a separate interactive terminal path, the Agent TUI guide, which makes the split between headless automation and terminal-first agents explicit.

The interesting bit is how narrow the packaging is. OpenRouter's guide screenshot describes a scaffold for headless agents, not a general chat app, while the launch post ties it directly to Bun and multi-model workflows. Even the first early test, via OpenRouter's repost of sensdlx, centers on a narrow Hermes worker producing strict JSON and X/Markdown artifacts.

OpenRouter's launch pitch is simple: generate a headless agent shell instead of wiring a coding agent from scratch.

The evidence points to three concrete claims:

• create-headless-agent is a new skill, per OpenRouter's announcement.
• The target is a customizable, multi-model headless CLI using Bun, again from the same post.
• The official explainer lives in OpenRouter's announcement post, linked from the launch tweet.

That makes this a harness story more than a model story. OpenRouter is packaging inference routing, tool use, and CLI setup into a reusable starting point.

The docs screenshot in OpenRouter's follow-up post gives the clearest product definition: "Build Your Own Headless Agent" under the Coding Agents guides.

From that page preview, the scaffold is aimed at:

• CLI tools
• API servers
• Automation pipelines

The same screenshot says the create-headless-agent skill uses @openrouter/agent for the inner loop. OpenRouter also calls out a separate "Build Your Own Agent TUI" path in the guide preview, which suggests the company is now splitting its agent templates by interface, headless for automation, TUI for interactive terminal work.

The only community signal in the evidence pool is still useful because it shows the template being tried on a constrained workload, not a generic demo.

According to that repost, sensdlx tested the skill for a narrow Hermes worker and got to "2 URLs in" with strict JSON plus X/Markdown artifacts. That is a small datapoint, but it reveals the shape of the intended output: structured artifacts and pipeline-friendly execution, not just conversational agent loops.