this one covers monday 19th to sunday 25th october, plus ten posts from the weekend of 17th and 18th october that went up after last week's issue was already frozen. most of the links this week are things you can use today: a map of the apps your Hermes Agent can work in, five Wingtips, three Autopilot prompts, the Meshy plugin and a full game asset kit built with it, a watch client, 25 new user stories, and the plugin catalog passing 750 entries. the local lane ran long too, with a voxel Mont-Saint-Michel series on the RTX 5090. thank you for reading.

this week's tips

five Hermes Wingtips, each run live before posting:

• #105: trust: untrusted: add trust: untrusted under an MCP server in config.yaml. tools the server marks read-only run as normal, and every other tool waits for your /approve.
• #106: Telegram messages sent while your agent was off: by default the gateway skips the Telegram messages that arrived while it was off. hermes config set platforms.telegram.extra.drop_pending_on_cold_boot false brings them in, in order, on the next start.
• #107: skill installs on a shared connection: the skills hub reads through the GitHub API, which allows 60 requests an hour per network address without a token. hermes config set GITHUB_TOKEN <your token> gives your Hermes Agent its own budget of 5,000 an hour.
• #108: hermes memory setup: one command walks you through adding a memory provider, for semantic search, knowledge graphs or automatic fact extraction. the built-in MEMORY.md and USER.md stay on the whole time.
• #109: hermes plugins search: hermes plugins search <term> matches a word against plugin names, descriptions and the tools each plugin adds, and the results end with the install line.

three Hermes Autopilot prompts, each with a copy-and-paste prompt in the first reply:

• #22: one agent or two: ask your Hermes Agent whether some of its work needs an agent of its own. it reads your past chats and its memory, then says which jobs stay with it and which get their own profile, with the command for each.
• #23: price watch: it checks the page once, shows you the price, asks what should trigger an alert, then schedules the check.
• #24: reminder until done: it asks what, when and where, then schedules a reminder that repeats until you reply.

things to know

longer reads, every line checked against the live docs before posting:

• the apps your Hermes Agent can work in today: a thread built from your answers to "which app next?", in seven groups: work and projects, team chat, email and calendar, on your Mac, infra, creative and support, and anything with its own MCP server. one more on top of Tonbi's connectors walkthrough: you can connect more than one account of the same app, for example Notion 'home' and Notion 'work'.
• sessions and context in Hermes Agent: what a session is and when /new makes sense, what carries over (MEMORY.md, USER.md, skills, and session_search over earlier conversations), why replies cost more tokens as a session grows and how auto compression at 50% plus /compress and /context handle it, how to start fresh in the Hermes Desktop Bot Chat, and /goal for a task that runs for hours.
• paying for models in Hermes Agent: Nous Portal next to your own subscriptions, what the Free plan runs and what paid plans add, the official Claude Subscription DirectSDK plugin for a Claude Pro or Max plan, and hermes fallback add for a backup provider when the main one hits a rate limit mid-task.
• five Hermes Agent features worth knowing about: ask Hermes to install plugins and skills for you, ask it for the Hermes Desktop tour, plugins that find the desktop app you already have, a live countdown on repeating tasks, and a prompt before unfinished code in a side session is thrown away. the first three are in v0.22.2, the last two land with the next release.

from Nous Research

• Beam 1.0 from Reflection, free on Nous Portal: open-weight, 501B total with 23B active per token, 262k context on the free version with tool calling built in, reasoning from low to max. hermes chat --provider nous -m reflection/beam-1.0-501b-a23b:free, or pick it in hermes model.
• Kimi K3.1 Pro on Nous Portal: one hermes model away in your Hermes Agent.
• the Meshy plugin: text, one photo or several photos to 3D, then texturing, remeshing, UV unwrapping, rigging and animation, all from a chat with your Hermes Agent. finished models download as GLB, FBX, OBJ, STL or USDZ at no extra credit cost, and the Hermes Desktop gallery keeps what each model was made from. Meshy is offering Hermes users 1,000 credits for $1 a month.
• Cua Spaces: when a task needs a computer, your agent creates a Space on your Mac, Windows or Linux box, does the work there and deletes it when done. one config line: hermes config set terminal.backend cua.
• the advanced-delegation plugin: I set it up with two routes on Nous Portal and gave it one prompt with two jobs. a quick summary of Python's heapq went to Claude Haiku 5.5 and came back in 3.0 s; a five-pirate logic puzzle went to Claude Opus 5.5 at high reasoning and came back solved in 9.3 s. the main agent picked the route for each job by itself, and the whole run is in the video.
• after the free trial, the next step is to login: a Nous Portal subscription gives your Hermes Agent hundreds of models, plus web search, image generation, browser automation and text-to-speech through the same account.
• the Microduck in Jensen's GTC Berlin finale: the keynote closed on a Hermes Agent demo training simulated Microducks on RTX Spark. mine is still on its way, and its walks, further down, are trained at home with Hermes Agent.
• Ornith-1.5-122B launched with Hermes Agent in its launch list: ollama launch hermes --model ornith-1.5:122b.

hermes desktop, on video

• the plugin catalog has passed 750 plugins: a 34 second tour of the catalog, memory, voice, messaging, browsing and more, built by the community.
• a full game asset kit with the Meshy plugin: about an hour and a half in one Hermes Desktop chat. a rigged duck knight hero and a rigged cultist duck enemy, each with four animations, nine props for a ruined level, 17 game-ready GLB files, and the knight as a 10 cm figure in STL and OBJ ready for a slicer. Hermes planned the kit with a credit cost per step, asked before every paid call, shrank the 4K textures to take the kit from 500 MB to 132 MB, and repaired the walls under 1 mm before the print check. 541 credits for the whole kit, 89 files in one project folder.
• hatch your own pet: type /hatch in any Hermes Desktop chat, describe it, and it hatches.

from the community

• Hermes Watch: an open-source voice client for the Amazfit Active 2. tap TALK, ask, and your Hermes Agent answers on the watch in text and out loud. Approve / Deny cards when Hermes asks before running something, long jobs keep running after you close the app with the result landing on your Telegram, and it runs on your own PC inside your Tailscale network.
• 25 new stories on the use cases page: a drainlaying business sending quotes and invoices from Telegram, one IT person at a 165-person travel company running about 80% of the daily work through Hermes, a food and beverage company on three Hermes profiles, and a mushroom grow tent on a Raspberry Pi. 395 real stories now. I also asked which kind you want more of, and the replies go into the next batch.
• 144 plugins added to the catalog in one week: the six I would install first: llmslim, hermes-memory-pane, advisor, agent-station, ux3-product-design-harness and hermes-pantheon, one line each in the post.
• Memory Pane: Adolan's plugin for a request we hear a lot: see exactly what your Hermes Agent saves, what it will remember and what it skips. he also shipped eighteen Hermes Desktop themes drawn as marble gods, the hermes-pantheon in the list above.
• vibeline: a timeline of your agent's sessions. each subagent branches off at the moment it was spawned, and you can click any step to see what happened there.
• plan-mode: your Hermes Agent reads the code, asks you a question or two where only you can decide, shows the full plan, revises it on request, and only then carries it out.
• Footprint: what your agent has left on your disk and what can go. each folder sorted by size, caches that get rebuilt flagged as worth a look, and the files your Bots wrote.
• Tabella: a chat becomes a real document. a trip budget comes back as an .xlsx with working formulas, a cover letter as a .docx or .pptx, and it reads Word, Excel, PowerPoint and PDF files back as text.
• Draft Cards: every email draft shows as a card in Hermes Desktop, so you see each one before it goes out and can ask for changes or edit it yourself.
• Jot: checklists and plans as notes beside the chat in Hermes Desktop, and your agent ticks off to-dos and adds new ones when you ask.
• wiki-ref: the Wikipedia text lands in your message next to your question. titles autocomplete as you type, you get the opening section or one named section, the source link comes with it, and it needs no API key.
• viajante: flights and hotels from your own machine, with no API keys and no account. it reads the public Google Flights and Google Hotels pages from your own IP address, returns prices, times and links, and books nothing.
• by2kb: give your agent a YouTube link, it transcribes it on your own machine and files a short abstract plus long notes in a folder.
• hermes-vcc-engine: compacts a long chat without asking the model to write the summary. it keeps every command with its exit code, the errors and the files it touched, and every line links back to the original message.

five videos from the community this week:

• 7 Hermes AI skills for a home lab: Proxmox, Docker, TrueNAS, Portainer, DNS and more, each with a first prompt you can run for your setup.
• Hermes Agent Telegram setup: how to reach your Hermes Agent from your phone, why the host has to stay awake, and what to check when the bot does not answer.
• the Anatomy skill: Hermes Agent explains an idea as an interactive machine figure.
• control a smart TV with a local Hermes Agent: five minutes on connecting Hermes Agent to your TV and clearing out the apps you never use.
• Hermes Agent for beginners: Wanderloots' walkthrough of a first local agent, the one to send to a newcomer.

local AI beyond Hermes

a long week on the RTX 5090, bench mornings, three new model launches and a voxel series:

• Microduck on a 25 mm step: 3,000 more rounds in the simulator on floors with low platforms, and it now stays on its feet in 11 of 15 tries, up from 5 of 15 at best.
• Coder390 against its base: a coder fine-tune of Qwen3.8-27B, same quant, thinking on. code and maths scores match with shorter reasoning, hard science comes out ahead, and instruction following (no commas, a set number of keywords, all capitals) slips more often than the base.
• Muse Glimmer 30B, speed: Meta's open-weight agent model on one RTX 5090 with room to spare, served at the 64K context Hermes Agent needs, and decode barely moves between an empty context and 32k. then the agent test: 10 of 10 short real tasks in Hermes Agent, not one malformed tool call.
• why your SSD is not extra RAM: llama.cpp memory-maps the model file, so what does not fit in RAM is read from the SSD when needed. a 79 GB model on a 59 GB RAM box took prompt processing from 710 tok/s to 260 tok/s once it got close to the limit.
• DeepSeek open-sourced DeepInfer: originally built on vLLM, NVIDIA SM100 and Huawei Ascend 950 only, one model for now, offline batch inference, batch-invariant and deterministic. it will not run on a 5090.
• vLLM's Humming kernels: quant kernels are the GPU code that does the maths on compressed weights, and Humming supports SM120, so a 5090 test is coming.
• EAGLE-4 on the 5090: the drafter guesses a small tree of next tokens and Qwen3.8 27B checks the whole tree in one pass. on the 4-bit model one request went from 57.6 to 175.9 tokens per second.
• running AI at home: bank statements, tax forms and government papers, worked on without anything leaving the house. the full rig behind it: RTX 5090 32 GB, Ryzen 5 9600, 64 GB of RAM, Ubuntu 26.04, llama.cpp's llama-server and Hermes Agent pointed at it.
• Spark-X2.6-4B in Hermes Agent: SparkLLM's 4B on-device model at the 64K context, with most of the card free. only 9 of its 36 layers read the whole context, the other 27 look at the last 512 tokens. first tool call came back clean.
• a voxel Mont-Saint-Michel, episode 1: one prompt and one hour of /goal in Hermes Agent for Qwen3.8 27B. the tide rises and falls on a one-minute loop, pilgrims on the causeway, sheep on the salt meadows, and keys for tide, rain and night. the v2 prompt finished in 31 minutes, and every video in the series now runs on it.
• Bonsai files, round 2: this year's AIME and HMMT, 63 problems, 32k thinking budget. Ternary-Bonsai-2-27B PQ2_0 at 7.2 GB scored 46, Qwen3.8-27B Q6_K at 22.9 GB scored 47, PTQ1_0 at 5.9 GB scored 43. question by question, PQ2_0 solved 6 that Q6_K missed and missed 7 that Q6_K solved.
• Flint on day one: 216 tokens per second on an empty context, 152 once 32K tokens are in. a small coding job inside Hermes Agent came back with the correct output.
• Qwen3.8 27B against Flint on the same voxel prompt: both finished on their own, Qwen in 31 minutes and Flint in 26. Qwen built most of the scene, Flint the mount and a rough abbey with no causeway. then the three-way with Muse Glimmer 30B: Muse finished first, in 21 minutes, with a cone-shaped mount and a night key that does nothing. one run each.
• Microduck's mixed-floor walk: down a 25 mm step and onto a hill without switching walks, plus slippery floors and shoves twice as hard as the default training. Hermes Agent queues the training runs, tests every walk and publishes each one on Hugging Face. both walks are free to download, links in the reply.

that's the week. if you are new here, the sessions-and-context and paying-for-models posts above answer the questions I get most, and the apps map tells you where your Hermes Agent can already work. see you next week.