I am very happy to present `hermes record`, a Hermes Labs plugin built by Nous Research.
it records what you do in your browser, creates a skill and Hermes can then repeat the task with whatever details you give it.
in this video, I record my pizza order, then use /learn to make a skill out of that recording, then ask my Hermes Agent to place a different order on its own :)
'hermes plugins install workflow-record'
AGI.
@mitsuharu_e thanks for flagging. when it doesn't start, does it show any message, or close straight away? and which OS are you on?
since you mostly use Telegram: on a terminal install, `hermes update` updates Hermes without opening the app.
one line to add it to your Hermes Agent:
`hermes skills install official/security/unbroker`
it finds the data broker and people-search sites that list your name, addresses and phone number, and files the removal requests for you 🕵️
tested Humming on my RTX 5090.
(1) on a 4-bit Qwen3.8-27B (NVFP4), switching vLLM to Humming took one request from 56.9 to 71.6 tokens per second, +26% over the kernel vLLM picks by default on this card.
(2) against Marlin, the older 4-bit kernel, it is a tie: 75.8 vs 77.1 on a dense int4 27B, 309 vs 292 on a 35B MoE.
(3) however, Marlin gets the same jump on that model (72.4). the default FP4 kernel is still the fastest at reading a long prompt.
how to try it: Humming already ships inside vLLM 0.32.0, no extra install. add `--linear-backend humming`.
don't upgrade humming-kernels on its own: 0.1.18 fails to load dense models in 0.32.0 on my box.
repo: https://t.co/jODo6IbtIs
another great win for local AI :)
quant kernels are the GPU code that does the maths on compressed weights and they're a big part of how fast a quantized model runs on your card.
Humming supports SM120 too, so I can test it on my 5090.
your Hermes Agent, on your wrist.
Hermes Watch is a new open-source community build for the Amazfit Active 2 watch: tap TALK, ask, and Hermes Agent answers on the watch, in text and out loud.
what it does:
- voice in, answers on your wrist, no phone in your hand
- reads answers out loud, with a Voice on/off button
- one ongoing conversation until you say "new chat"
- Approve / Deny cards when Hermes asks before running something, one request at a time
- long jobs keep running after you close the app, and the result lands on your Telegram
- texts, car commands and a PC restart only run after you tap Yes
- runs on your own PC, reachable only inside your Tailscale network
what you can do with it:
- ask a quick question while you walk or cook
- approve the code change your agent wants to make, away from your desk
- start a build and get pinged when it's done
- check on your PC with a quick health check
opensource code below