witcheer✓

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.

five new Hermes Agent features worth knowing about: (1) see what a paid MCP tool costs before it runs (#142470) an MCP server entry can declare what each tool costs and a spend limit per chat. the approval prompt shows the cost and the chat's running total, and a call that would pass the limit asks first. (2) the classic CLI has a new default look (#139241) every line in the scrollback starts with a single glyph, and the status line sits under the input box. (3) long MCP tool calls show their progress (#141985) when an MCP server reports progress, Hermes Desktop shows the percentage and message next to the running tool's timer. the CLI spinner shows the same text. (4) pick the model your subagents use from a dropdown (#141312) Hermes Desktop settings now have separate provider and model dropdowns for subagents. (5) drag a session to pin it (#142672) in the Hermes Desktop sidebar, drop a session on the Pinned section to pin it, and drag it back to the session list to unpin it.

docs for each: (1) (2) (3) (4)

when does "0 9 1 * 1" run? if you said the first Monday of the month, it runs on the 1st and on every Monday. crontab-ref is a community plugin that gives your Hermes Agent the plain-words reading of any cron expression in your message: what each field means, the next five run times, and whether /cron will take it. no API key needed

built by olric3015 on GitHub, thank you for sharing it. catalog: hermes plugins install crontab-ref hermes plugins enable crontab-ref

your Hermes Agent isn't locked to one model provider. `/model` switches it mid-chat between the providers you've set up: Nous Portal, OpenRouter, OpenAI, Anthropic, or your own llama.cpp server. your conversation history, memory and skills carry over, whichever one runs it. to add a new provider, run `hermes model` in your terminal.

your Hermes Agent can message you first: when mail arrives, a file is shared or a build finishes. hookspool is an open-source community project: a local inbox in front of your Hermes webhook routes that keeps every event on disk until your Hermes Agent takes it. > it saves each webhook to disk before it answers the sender > it holds events while your gateway restarts and delivers them when it is back > it turns a burst of events into one wake-up > bobbins, one plugin per service: Gmail, Google Calendar, Google Drive, GitHub, Trello, Kaneo, Zoom and Plaid > one general bobbin for any service that sends signed JSON > setup steps in every bobbin, one for you and one for your agent > it works with the Hermes webhook routes you already have > plain Python and one SQLite file, MIT licence what you can do with it: > get a message on Telegram, Discord or Slack when a new email arrives, with a draft reply from your agent > get a heads-up when someone shares a file with you in Google Drive > get one message after a burst of CI runs > ask your agent for a meeting brief when a new meeting lands on your Google Calendar

built by jaredzimmerman on GitHub and reddit, thank you for sharing it: his post with the story behind it:

Kimi K3.1 Pro is on Nous Portal, so I gave it the Mont-Saint-Michel voxel prompt from my series, with one hour of /goal in Hermes Agent. it built the abbey with its golden statue, the village inside its ramparts, pilgrims on the causeway, sheep in the salt meadows, a tide coming in, a rain shower and a night with lit windows. it used the whole hour. its last edit to the page came at 44 minutes, and it spent the rest testing the page in a browser and checking its own screenshots. the run cost $3.38.

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:

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.

good way to start a Sunday 👀

the community added 144 plugins to the Hermes plugin catalog this week. the six I'd install first: > llmslim: compacts older tool output on your machine, with no model call to summarise it, so prompt caching keeps working > hermes-memory-pane: shows what your Hermes Agent saved to or forgot from memory in the current chat, and lets you edit or undo it (Hermes Desktop) > advisor: a second model reviews every turn and sends its findings back as advice > agent-station: a read-only Hermes Desktop page for the subagents your Hermes Agent starts, with a replay of file edits > ux3-product-design-harness: lets your Hermes Agent review a product decision before you build it and give you one verdict > hermes-pantheon: eighteen Hermes Desktop themes drawn as marble gods, switched with /skin <name> thanks to everyone building and sharing these!

credit on GitHub: Thanatos9404 (llmslim), Adolanium (hermes-memory-pane, hermes-pantheon), ljubomirj (advisor), thealps01-netizen (agent-station), cis2042 (ux3-product-design-harness). advisor is a port of pi-omplike-advisor by pasky. the catalog, with every plugin's page, tier and listed date: install any of them with: hermes plugins install <name>

planning a long weekend away usually starts with a pile of flight and hotel tabs. viajante is a community plugin for Hermes Agent that searches flights and hotels from your own machine, with no API keys and no account. everything it covers: - reads the public Google Flights and Google Hotels pages from your own IP address - returns prices, flight times and links - gives your agent 22 tools plus a skill that explains how to read the results - books nothing and stores no credentials what you can do with it: - ask for a round trip with your dates and currency, and get nonstop options with flight numbers and times - ask about both Paris airports and have it pick one flight and one hotel, with its reasons - find hotels for the same dates, priced for the whole stay, with a minimum rating

built by felipebasurto on GitHub, thank you for sharing it!

RT @NousResearch: Hermes Agent gives you control

Person who is reverse engineering his tv gets banned from OpenAI You aren't allowed to stop your tv from spying on your family. You have to ask for permission Or, maybe, maybe we accelerate

my Hermes Agent now sends each subagent to the model that fits the job. I set up the new advanced-delegation plugin from Nous Research with two routes on Nous Portal, then gave it one prompt with two jobs: - a quick summary of Python's heapq went to Claude Haiku 5.5 and came back in 3.0s - a five-pirate logic puzzle went to Claude Opus 5.5 at high reasoning and came back solved in 9.3s the main agent picked the route for each job by itself. the whole run is in the video, setup in the first comment.

hermes plugins install advanced-delegation needs Hermes 0.22.5 or newer, then `hermes delegation config` to set your routes

I gave Qwen3.8 27B, Flint and Muse Glimmer 30B the same Mont-Saint-Michel voxel prompt on my RTX 5090, with one hour of /goal in Hermes Agent each: >Qwen built most of the scene: the abbey, the tide and a night with lit windows. >Flint built the mount and a rough abbey, with no causeway and a night too dark to see. >Muse finished first, in 21 minutes, with a cone-shaped mount, a spire and a night key that does nothing. one run each. Flint ran at its maker's settings, Muse on defaults since its GGUF ships none (its maker-settings rerun did worse).

full method &amp; per-model breakdown in the Grimoire 📖 🔧 🤗