last week's issue went out this afternoon, so this one is monday 14th to sunday 20th, plus four sunday-13th posts that landed after that draft was already frozen: wingtip #71, a throwaway-profile habit, a spec-cache bench, and the Bangkok meetup. Hermes Business landed for teams, the plugin catalog opened, and I posted three more Autopilot prompts. eight Wingtips, five community builds, and on the 5090 a think-on board.

this week's tips

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

• #71: hermes -z --resume: hermes -z is the scripted one-shot (prompt in, final answer out, nothing else on stdout). --resume latest on the next -z call replays the earlier turns so the second prompt builds on the first, same session id.
• #72: skills.create_dir: new skills normally land in the profile-local skills folder. this key redirects skill_manage create to a named folder, made on first write; skills there join the index, hermes skills list and the slash commands. existing skills stay where they are.
• #73: gateway.message_timestamps: off by default. when on, Hermes prepends the send time to every user message in the model's context only. the visible chat and the on-disk transcripts stay clean; time is stored as metadata either way, so turning it on later also surfaces past send-times.
• #74: provider_routing: on OpenRouter one model name is served by several hosts. ignore: under provider_routing in config.yaml keeps that model off a listed host even when it is the cheapest; only: and order: are there too. edited in the file, not via hermes config set. on Nous Portal there is nothing to set.
• #75: reasoning_effort for side tasks: compaction, session title, vision, the approval classifier and the curator each take their own thinking level (none through ultra), separate from agent.reasoning_effort on the main chat.
• #76: display.vim_mode: vi keys in the Hermes CLI input box. off by default, config-only, read at startup. Esc is NORMAL, i is INSERT, and the live label sits on the right of the status bar. CLI only.
• #77: display.bell_on_complete and display.bell_on_prompt: a terminal bell when a turn finishes, and when a blocking prompt opens (clarify, approval, sudo password). both also emit an OSC 9 desktop notification in Ghostty, iTerm2, Kitty and WezTerm. both off by default.
• #78: approvals.smart_policy: your own rules for the smart-approval reviewer, appended to its system prompt. the reviewer answers APPROVE, DENY or ESCALATE on flagged terminal commands. empty by default; approvals.mode: smart is the default mode.

two from my own setup:

• a throwaway profile for testing prompts: hermes profile create test --clone copies config, .env, SOUL.md, skills and the two memory files, and leaves sessions and cron behind. hermes -p test runs on the copy; hermes profile delete test throws it away.
• hermes config get <key> before you write about a setting: get, set, unset. I run the round on a throwaway profile so the default comes back clean.

shipped this week

two launches, two explainers, and a few things that change how you sit in front of the product:

• Hermes Business, what a team gets: the day-two walk through the portal FAQ. data handling, running on your own infra, a shared skill library, a separate team balance with per-member caps, roles and channels, and how models are billed. Enterprise is the on-prem / your-cloud sibling with SSO.
• the plugin catalog: plugins were files on disk already; this is the reviewed list inside the product. vetting, updates after review, new entries without updating Hermes, installed is not enabled, your own repo by URL, and how to submit (one pull request, one yaml).
• Refactoring Hermes with 1,393 agents: the Nous blog on a large internal run. the levers in it are in your install today: /goal, delegate_task with worktrees, and skills the agent writes for itself.
• Teknium on this month's focus: his post, linked rather than restated.
• one composer, several sessions: Brooklyn's Desktop demo. tabs for parallel sessions, and the session under your cursor is the one that hears you.
• the HermesWatcher blocker results: I read 150+ replies to his "biggest thing stopping you" thread and wrote back what you named, with two product beats (local models on Desktop, browser.real_profile_pin) and the docs links in the first reply.
• which surface I use for what: terminal, Desktop, Discord from the phone with /handoff discord, Hermes Cloud as the always-on backend, and hermes profile create when I want a second agent.

hermes autopilot

prompts you paste into your agent, which then does the work. #4 and #5 keep the prompt in the first reply; #6 puts the prompt in the post itself:

• #4: the Monday review that writes itself: it reads session history, writes a ten-line review, shows you the cron job, and creates it only on your yes.
• #5: write the AGENTS.md for this repo: it drafts the file, shows you the whole thing, and writes nothing until you say yes. Hermes's own protected-file gate on AGENTS.md still fires.
• #6: a daily brief on the topics you name: a sourced brief, a link on every line, cron created only on the yes. the full prompt is in the post.
• the Autopilot index: the prompts on one hermes-recipes page. the page had four the day I posted it; #5 and #6 went up after.

builders worth a look

• mykeura's Minimalist Themes: eighteen light, warm palettes for Hermes Desktop. a drop-in desktop plugin, no core patching.
• hermes-project-stewardship: a plugin that puts a team of Hermes profiles in charge of a repo. it checks git, tests and CI, proposes work on the Kanban board, and waits for approval. autonomy from read-only through to pull requests.
• a family voice hub on the fridge: shantanugoel's household Hermes, voice in the kitchen. 32 seconds of it running.
• Hermes-Portal: a read-only dashboard over a live install: skills, conversations, cron, spend, memory, plugins, health. Python stdlib, no writes.
• hermes-lcm: a plugin context engine for long conversations. every message stays in a plugin-local SQLite store, older turns fold into a summary DAG, and recall tools page exact lines back. it fills the documented context.engine seam.
• Tonbi's MCP Command Center tour: per-profile servers, the Capabilities scope selector, Desktop hosting the OAuth callback for a remote backend, and the once-a-day Sign in nudge when a token expires.
• AAIF NYC, 15 september: Agentic AI Night with Nous Research, Manhattan, Karan on the panel.
• Bangkok, friday 25 september: free, 18:00 ICT at Cleverse, hosted by Jon Komet. Nous-sponsored, same shape as Berlin. if you want to host one where you live, tell us.

local AI beyond Hermes

two bench posts on the RTX 5090 this stretch, both llama.cpp:

• spec decode and prefix cache on Qwen3.8-27B Q6_K: prompt-cache TTFT 1.004 s to 0.069 s on a 2,954-token prompt, and 3.814 s to 0.086 s at 11,570 tokens. MTP draft head n=2 took decode from 61 tok/s to 109 prose / 113 chat / 139 code / 144 repetitive. stacked, 4.25 s to 0.30 s. 41% of answers differ under MTP; the flips are the model's own second choice. (source: that post, 13 september)
• the second-tier think-on 16k board: ten local models, thinking on, four tasks (IFEval, MATH-500, HumanEval+, MBPP+), 1,583 items each. a three-way tie at the top, and a 12 GB file is in it: Gemma 4 31B Q4_0 90.1, Qwen3.8-27B Q6_K 89.7, Qwen3.8-27B UD-IQ3_XXS 89.5. (source: that post, 20 september)

the data lives in the rtx-5090-benchmarks dataset on HF and the llm-bench-rig repo.

I also started posting the Microduck work. sim only, MuJoCo on the 5090, nothing on a real robot yet:

• first stand: Pollen's walking policy, 12 s clip.
• iteration 0 against iteration 4000: first own velocity policy, 4096 envs, 58 min, side by side.
• sit and stand on command: SitStand PPO, 15,000 iterations in 4 h 32 min, mean reward 1.71 to 133.43. stand 116 mm, sit 59 mm trunk height.

that's the week. Bangkok is friday. if you have not opened the plugin catalog yet, that is the one to click. see you next sunday.