All stories
Voice Gate Ecosystem Community reports

Voice Gate gives Hermes one more chance to cut the AI writing tells

The new community plugin checks changed Markdown before a turn ends, while a personal edition can tune its rules to your writing habits.

By AITrending2 min read
Sources (5)

Why now Apostol Apostolov published Voice Gate 1.0.0 at 20:07 UTC on November 4, with tagged public and personal editions for Hermes Agent 0.22.0 or newer.
Voice Gate release artwork says it finds AI tells in Markdown and asks for changes before a turn ends, beside a quill and printing press.
Author-published public-edition release artwork, not a product screenshot or an independently reproduced writing result. Source: Apostol Apostolov

Voice Gate turns complaints about AI-sounding prose into a revision prompt inside Hermes Agent. The community plugin’s November 4 release checks the Markdown a turn changed and asks the agent for another pass before it finishes.

It targets recognizable English habits: verdict-style openings, closing offers of more help, repeated conclusion signposts, bold pseudo-headings and clusters of buzzwords. Thresholds matter: one bold heading or an isolated word is not automatically a finding. These are writing heuristics, not a detector that establishes who wrote a document.

A hook for the prose you just changed

Version 1.0.0 requires Hermes Agent 0.22.0 or newer. Users install the public edition from the repository’s public/voice-gate directory with Hermes’s plugin installer, then enable voice-gate. It is a plugin for the agent harness, not a model provider or a standalone editor; the repository carries an MIT license.

The public implementation registers a pre_verify hook. It reads the changed paths Hermes supplies, filters for .md, .markdown and .mdx files, skips common build and dependency directories, and ignores documents below its default 200-word floor. It returns a message requesting revision rather than editing the files itself. Unreadable files and caught errors return without blocking the turn.

Warn mode asks once. Improve mode can keep asking within the host’s nudge budget; skip disables the gate. The bundled public checks use local Python rules with no extra package or external model call in the inspected hook. Any subsequent agent revision still uses the user’s configured Hermes model and its normal usage terms.

Make the rules personal

The personal edition adds YAML or JSON profiles under HERMES_HOME/voice-profiles, plus tools to check text, inspect or validate a profile, and build a draft from a corpus of writing. Its learning tool can read chat exports, mail or documents and save a draft profile when requested, so that edition needs access to the samples the user supplies. That is broader than the automatic changed-file hook.

The practical appeal is specific: catch repeated prose habits while an agent still has the task open. Whether the nudges improve a particular person’s writing remains untested by AITrending; the author’s rule list and workflow are documented release claims.

Related reading