this one covers monday 21st to sunday 27th september. two more tags went out, v0.21.4 and v0.21.5, and with them Bot Screen, Fast Search, Simple mode in Hermes Desktop and the Claude subscription plugin. I published Flightplan #4, a seven-day plan for your first week with Hermes Agent. six Wingtips, three Autopilot prompts, two Logbooks, and the Microduck's whole skill tree went up on Hugging Face, with Hermes Agent driving the duck from one typed sentence by thursday.

this week's tips

six Hermes Wingtips this week, each checked against the live docs before posting:

• #79: agent.disabled_toolsets: each platform keeps its own toolset list in hermes tools. this key sits above all of them: a toolset named here is removed everywhere at once, CLI and every gateway platform, applied after the per-platform lists. web takes web_search and web_extract out of every session; memory removes the memory tools and the memory guidance from the prompt.
• #80: skills.auto_load: a list of skills fully loaded at the start of every new session, on the CLI, every gateway platform and cron jobs. resolved once when the prompt is first built, so the prompt stays cache-stable; an edit to the skill shows in the next session. hermes -s <skill> does the same for one launch. per profile, empty by default.
• #81: unauthorized_dm_behavior: when someone you have not approved messages your Hermes bot, it answers with a one-time pairing code by default. set it to decline and the bot sends one short polite decline, then stays silent toward that sender for a day.
• #82: agent.session_stall_timeout: the "Agent session appears stalled" message on a gateway bot is a watchdog and never stops the turn. it speaks up once when you have sent a follow-up and the agent has shown no activity for 5 minutes; then it is your call, /new, /stop or keep waiting. raise it for long jobs, or 0 to turn it off.
• #83: approvals.denial_breaker_threshold: when the smart-approval reviewer denies a command, the agent can try a variation. this key counts denials in a row per session; at 3 the deny message tells the agent to stop, report what was blocked, and ask you to run it or /approve it. lower it to stop sooner, 0 turns it off.
• #84: mcp.discovery_concurrency: a cap on how many MCP servers connect at the same time when Hermes Agent starts, each local one being its own process. raise it to bring a long list up sooner, lower it on a small machine.

the long read

• Flightplan #4: your first week with Hermes Agent: seven days, one outcome a day, about twenty minutes each. install it and have one real conversation, let it learn how you work, turn something you repeat into a skill, reach it from your phone, give it a job that runs without you, bring it into a project, then look back and tidy up. the claims were re-checked against the live docs after v0.21.5 before it went out.

shipped this week

two tags, two Logbooks, and the week's launches with the questions asked under them answered:

• Logbook #9: v0.21.2 and v0.21.3: five things from the two tags. the agent can log you into websites, with the password filled from 1Password, Bitwarden or its own encrypted vault and never seen by the model; a check before an expensive model switch in a long chat; cron jobs that fire from a webhook event with the schedule kept as fallback; long pastes in Hermes Desktop become an attachment chip; and /queue in the CLI lets you edit, remove or reorder waiting prompts.
• Logbook #10: v0.21.4 and v0.21.5: a workspace picker on the web dashboard chat so a new chat starts in the project you choose; Claude chats keep their cache for an hour with prompt caching on auto; /branch opens a new thread on Discord, Telegram, Slack and Matrix; Hermes Desktop offers to retry when a provider's usage limit lifts, with a countdown you can cancel; and background commands get their own row in the CLI and TUI live dock.
• Bot Screen, five answers: it runs on a Linux gateway host and each bot gets its own desktop there, not yours. while you hold control its computer-use and browser tools are refused, screen captures included. at a login, 2FA, CAPTCHA or payment it tells you in chat and ends its turn; you do the step and hand back. plan on 1.1 to 1.5 GB of memory per open screen with a browser. off by default, started from the Screen pane in Hermes Desktop or hermes computer-use screen start on the host.
• Fast Search, four answers: if your web search already runs through your Nous Portal subscription it switches over with the update; otherwise hermes tools moves web to the Nous route. same web_search tool, so the agent sees nothing new. searches go through the Nous Tool Gateway on your Portal login, no Perplexity account or key on your side. your agent gets ranked results with snippets, then reads the pages it picks. your own key or SearXNG keep working as before.
• Simple mode in Hermes Desktop: the chat, your sessions, Capabilities and Messaging, and nothing else on screen. Hermes still does everything it did; Advanced brings the full workspace back. this came out of the simplicity intake I ran last month, where a Desktop for people who are not developers was one of the clearest asks.
• the Claude subscription plugin, explained: your Hermes Agent keeps its own loop, tools, approvals, memory and compaction; the official Claude Code CLI acts as the model client for each request and nothing more. authentication stays with the CLI, and the plugin never opens or copies its credential files. quota and extra-usage settings stay in your account. Linux, macOS and Windows, v0.21.4 or newer, in hermes model and the Desktop and TUI pickers. an official Nous plugin, marked experimental, MIT on GitHub.
• the plugin catalog pages: before you install, each page shows the exact commit the maintainers reviewed (the one hermes plugins install <name> checks out), the tools, hooks and environment variables the plugin declares, the README and screenshots from that commit, the Desktop install button next to the CLI command, and a shelf of everything else by the same author.
• Capabilities, then Plugins, in 44 seconds: one row per plugin in Hermes Desktop, a badge for whether it extends the app, the agent or both, and two switches per row. in the video I install hermes-rss from the catalog and point it at the Hermes Agent releases feed.
• compaction on 8k to 32k local models: when the context fills, Hermes summarises the older part and keeps the recent part word for word. that protected tail is now capped at about 20% of the model's window: 2,116 tokens kept on an 8k window, 4,246 on 16k, 7,441 on 32k, unchanged from 128k up. protect_last_n still sets a minimum.
• Hermes Desktop in 9 languages: English, Simplified and Traditional Chinese, Japanese, Arabic, Russian, and now French, German and Spanish. Settings, Appearance, Language.
• the Kanban task modal: Tony Simons showed the new task view in Hermes Desktop. one thing he did not show: on a running task, the comment box goes straight to the worker, so you steer it from the same window.

hermes autopilot

prompts you paste into your agent, which then does the work. the prompt sits in the first reply, the proof run in the second:

• #7: a skills inventory your agent writes for you: skills are folders with a SKILL.md, so the agent opens every one, rewords each description, checks the body against it, and looks for two skills doing the same job. the result is skills-index.md.
• #8: your agent audits the files that shape it: SOUL.md, its two memory files and your project's AGENTS.md, each line checked against your machine and the project files.
• #9: your agent writes its own manual: model, provider, platforms, cron jobs, skills, where memory and sessions live, and how to back it up. each line checked on your machine, keys shown as set or not set.

from the community

• what you told me about Hermes for more than one person: the results of last week's question. most of you run one gateway that several people talk to; one profile per person on the same box is the other common shape; some go to a container or VM per person. the thing you want most is per-user permissions inside one shared bot, and several of you built that layer yourselves already.
• 326 use cases on the docs page: 64 added last month, 45 from reddit and 19 from X, every one linked to the real post. mostly dev workflows, then business ops, creative work and cost-cutting setups.
• Jon Komet's Hermes Browser Extension: a side panel that puts your own Hermes Agent inside Chrome, Edge or any Chromium browser. you stay in control, Hermes sees the tab you pick, and on a local or self-hosted gateway it uses the models, tools, skills, memory and MCP servers you already set up.
• a printed daily paper: aidaniil has Hermes print a one-page newspaper each day. the four pieces are all in your Hermes Agent already: a cron job, the terminal over SSH to a Mac with a printer, the google-workspace skill and MCP servers for inputs, and lp for the PDF.
• Tonbi's plugin catalog tour: a six-minute walk through installing from the catalog in Hermes Desktop, with the four facts that matter: every entry came in through a reviewed pull request, an entry pins one exact commit, hermes plugins install checks out that commit, and installed is not enabled.
• Tonbi's Fast Search how-to: under two minutes, the whole process on Nous Portal.
• Circuit Berlin power users: the recap from the Berlin meetup. thanks to Circuit and to everyone who showed up with their setups.

local AI beyond Hermes

no bench run this week; the 5090 went to the Microduck. still simulation only, MuJoCo on the card, nothing on a real robot yet:

• level 6, a forward roll: 10,000 iterations, 4096 simulated ducks, 2 h 53 min. half the attempts start mid-roll so the duck practises getting back up first. the reward curve went flat at iteration 2,000.
• the whole skill tree on Hugging Face: six levels, 26 h of card time on one RTX 5090, five of them installable on a real Microduck with one command. the training log sits next to them as a dataset: reward curves, the queue item behind each level, take logs and clips.
• level 7, a kick: 10,000 iterations, 2 h 41 min. the policy never sees the ball, because the real robot has no ball sensor; the ball is placed at the toe every attempt and the duck kicks whatever is there.
• Hermes Agent drives the duck: one typed sentence, "make the duck walk forward for five seconds then sit". Hermes Agent checked the duck was sitting, stood it up, walked it 0.58 m and sat it back down, every command its own pick. running on a local Qwen3.8-27B.

that's the week. if you are new, Flightplan #4 is the place to start. see you next week.