Voice skills
A voice skill is an agent skill written to be heard: it answers in a few spoken sentences. Install them from Vux's settings, or ask your agent: "install the weather skill." Anyone can write one.
herdr-voice
herdr-voice runs coding agents in Herdr by voice: Claude Code, Codex, omp and any other agent Herdr supports. It needs Herdr on the computer your agent runs on, Hermes or Vux's own.
- "Create a new Claude session for vux."
- "Tell the vux Claude to fix the failing test, and let me know when it's done."
- "What's waiting on me?"
- "Where did we do the swipe fix?"
- "Merge it once CI passes."
A new session opens in the repo's Herdr workspace, named after its task, so you can refer to it by what it's doing. Your agent doesn't sit waiting while a coding agent works: you keep talking, and it tells you when the agent is done or stuck, with one report when you started several at once. "What's waiting on me?" goes through every agent and tells you which have a question for you and which finished with something you haven't heard.
With Hermes, agents you started yourself at your desk get the same treatment through herdr-voice's desk-alerts Herdr plugin: they tell you when they finish or get stuck while you're away, and you answer their questions and permission prompts from a card in Vux, or out loud.
Write your own
The user is listening, often away from the screen. Have the skill tell the agent to answer in two to four sentences, say what changed and what's left, and never read out terminal output, file contents or long lists. Speech-to-text mishears names, so list the likely mishearings of yours. herdr-voice is a worked example.
Describe it to Vux
Any agent skill works in Vux, Hermes' and Vux's own agent's alike. To give Vux more to show, add a metadata.vux block to your SKILL.md frontmatter; every field is optional. Agents ignore it; Vux reads it.
---
name: weather
description: "Weather and forecasts, in a sentence or two"
metadata:
vux:
voice: true
summary: "Says the forecast in a sentence or two."
requires:
command: curl
phrases:
- "what's the weather tomorrow?"
- "do I need an umbrella?"
routing: "For weather and forecasts, load the weather skill."
---
voicetruelists the skill under Voice skills in Vux however it was installed.summary- One sentence shown under the skill's name, up to 120 characters. Without it, Vux shows the
description. requires.command- A command the skill needs on the computer your agent runs on. Vux shows it, and the agent says so if it's missing after installing.
requires.capabilities- What the agent must be able to do:
background_tasks(hand work to a background agent withdelegate_taskand report when it's done) ormemory(remember across conversations). Vux says on the skill when the computer's agent can't, before it's installed and after. phrases- Up to six things to say, shown as examples.
routing- When the agent should use the skill, up to 300 characters. While the skill is installed, Vux adds this sentence to every turn, marked as coming from your skill, and shows it in the skill's details. Hermes' skill index sees only the first 60 characters of
description, so this is how a skill wins over a built-in one.
Hermes shows the model only the first 60 characters of description, so put what matters first.
List it
Vux's list of voice skills is the public repo tensor-systems/voice-skills. To add yours, open a pull request with an entry pointing at your skill's folder on GitHub. A check makes sure your SKILL.md can be fetched and its name matches. Skills Vux's maintainers have looked over are labelled reviewed; the rest are labelled community.
Listing is optional. Anyone can install a skill by its GitHub path, from Settings → Voice skills in Vux or by giving your agent the path. Hermes installs it as a community skill and scans it first; Vux's agent doesn't scan skills, so install only ones you trust.