Hermes LLM agent bid for NonGNU ELPA sparks SaaSS fight

Thanos Apollo has proposed hermes, an Emacs frontend for the Hermes Agent harness, for NonGNU ELPA. The package gives Emacs a dashboard and chat UI over local or remote agent instances, with streamed replies, approval prompts, session controls, and browsers for tasks, skills, and related services.

The submission drew a sharp free-software objection. Jean Louis argued that Hermes, though libre, is built around remote proprietary language-model APIs: its installer and defaults push a Nous Research portal subscription, and most users take that path because local inference is harder. Packaging a polished Emacs client, he said, normalizes Service as a Software Substitute rather than teaching GNU principles, even if a local backend is technically possible.

Maintainers disagreed on the bar. Eli Zaretskii noted that the setup wizard states users may press Ctrl+C at any time to leave configuration, and that exiting skips portal registration once the software is already installed. Philip Kaludercic stressed that NonGNU ELPA requires full function on a free-only system, not that every package refuse proprietary network services or preach GNU ideology. Apollo added that any compatible endpoint works, including self-hosted options such as Ollama.

Richard Stallman pressed for basics: what advantage an agent has inside Emacs over ordinary commands and Lisp, what "conversation" means, how the system learns, and whether it depends on an LLM. Apollo replied that Hermes is an agent loop around a model. It needs some LLM at runtime but not a commercial one, and its value is delegating multi-step goals (inspect, test, edit, re-test) under user approval, not replacing simple edits.

Tomas sided with the critics on UX, calling the installer's ambiguous exit wording a dark pattern that shepherds users while staying formally free. The dispute leaves open how Emacs archives should treat free clients whose easiest path still centers proprietary cloud AI.