Subclass the OpenAI-compatible gateway so third-party chat clients keep markdown and slash commands
A drop-in plugin runs a second gateway port that renders full markdown and intercepts /help, /status and the rest instead of forwarding them to the model.
So I built a small drop-in plugin that fixes both. It's a separate gateway adapter (subclasses the official one, doesn't fork it), runs on its own port
— @dworfd, 2026-05-08, via the public Nous Discord archive.
Third-party OpenAI-compatible clients such as BoltAI reach Hermes through the API server, where the contributor found markdown flattened to plain text and slash commands passed to the model as ordinary messages. Rather than fork the gateway, the plugin subclasses the official adapter, listens on a second port, renders markdown and handles slash commands locally, and preserves streaming. The client is pointed at the new port and otherwise configured as the official guide describes.
Limitations: A third-party plugin tested by its author with BoltAI; other OpenAI-compatible clients are stated as untested. It depends on the official gateway class it subclasses, so an upstream change could break it. Review the code before installing.
Documentation check: this specific community technique was not found in the official English reference snapshot (1f2b9e0) or the cached prompt guides. Related built-in features are documented. This is a bounded source review, not a claim about every document or independent execution testing.