it has been a month since the last issue, so this one is a catch-up. the short version: twenty-seven new Wingtips, two releases (Quicksilver and Herald), Hermes Desktop grew a HUD mode and an in-app browser, and Browser Use became the default way Hermes Agent drives the web. the local-ai corner is at the bottom as usual.

this month's tips

twenty-seven Hermes Wingtips shipped since the last issue. each one was verified against the live docs, and most of them I ran end to end on my own boxes before posting:

• #20: prune cuts the db down, export takes a session out of it: hermes sessions export writes a session as md, html or jsonl; --redact scrubs keys and tokens before anything leaves the box.
• #21: credential pools: several keys for one provider, rotated on rate limit; fallback providers are the separate, cross-provider mechanism.
• #22: steer a task without stopping it: /steer injects a mid-run note that lands after the current tool call; /queue holds a message for the next turn instead.
• #23: local model prints the tool call instead of running it: tool calling is switched on server-side; vLLM and SGLang take a --tool-call-parser, llama.cpp wants --jinja, and the parser has to match the model family.
• #24: your bot did not forget you: gateway sessions keep history across restarts by default; auto-reset (idle, daily or both) is opt-in, and memories are saved before any reset runs.
• #25: /goal completion contracts: a vague goal gets vague judging; tell the judge what done means with verify: / constraints: / stop when: lines, or let /goal draft write the contract for you.
• #26: the web page summariser burns your main model: by default it reuses your chat model; two auxiliary config lines route it to a cheap flash-class model instead.
• #27: /background runs a second task in parallel: a separate agent session with its own conversation; it starts clean, so write the prompt self-contained.
• #28: cron jobs and workdir: a cron job wakes up detached from any repo; --workdir /abs/path loads that directory's AGENTS.md and points every file tool at it.
• #29: the local context floor: Ollama defaults as low as 4,096 tokens of context while agent work wants 64,000; the fix is server-side (OLLAMA_CONTEXT_LENGTH, --max-model-len, -c).
• #30: back up your agent before you need to: hermes backup writes one zip of config, auth, memories, skills, sessions and cron; hermes import puts it all back.
• #31: see what every message costs: hermes prompt-size prints the fixed per-call budget of a fresh session, fully offline.
• #32: read the last 30 days off your own box: hermes insights parses the local session db into tokens, models, platforms and top tools.
• #33: checkpoints: opt-in undo for agent file changes; snapshots land before every write, /rollback lists and restores.
• #34: hermes proxy: serve your Nous Portal subscription to any OpenAI-compatible app; the app uses a placeholder key and the proxy attaches the real credential per request.
• #35: hermes bundles: group several skills under one slash command; a bundle is a small yaml file, so it can live in a dotfiles repo.
• #36: hermes import-agent: bring a Claude Code or Codex CLI setup across in one command; instructions become memories, skills become skills, and secrets are never imported.
• #37: hermes project: named multi-folder workspaces; Desktop groups sessions by project and bind-board links a kanban board to it.
• #38: hermes doctor: the built-in health check; one pass/fail report across config, auth, python env and skills, and --fix attempts the repairs.
• #39: the missing provider prefix: a model id without its vendor/ prefix fails with a 404 error; write vendor/model always, and hermes model picks it right for you.
• #40: scoped sudoers: give your local agent the two commands it needs, not root.
• #41: approvals suggest: mine your own approval history into an allowlist proposal; it proposes, you pick, and destructive command classes are never proposed.
• #42: the compaction-summary model: compaction summaries run on your main chat model by default; you can move that job to a flash-class model with two config lines.
• #43: Telegram topics or a second agent: /topic gives one bot parallel sessions with one agent behind them; hermes profile create gives you a genuinely separate agent with its own bot.
• #44: curator backup and rollback: the curator snapshots your skills library before every pass, and rollback restores one, so the undo is itself undoable.
• #45: claw migrate, the preview-first move: hermes claw migrate --dry-run previews an OpenClaw move without touching disk; secrets stay behind by default and a restore-point zip is written first.
• #46: Browser Use: as of the latest release it is the default browser mode; the card in the post carries the numbers and the opt-out.

shipped this month

the longer pieces, all tested on my own hardware before they went out:

• YOLO mode, properly explained: what /yolo skips, the three ways to start it, and the hardline floor of dangerous commands that stays on no matter what.
• Hermes Logbook: between v0.19.0 and the next tag: five features that merged before the release notes could catch up, streaming TTS and mid-turn redirect among them.
• SSH remote backend for Hermes Desktop: the Desktop UI on your laptop, the agent and its tools on a box you own, in the VS Code Remote-SSH shape.
• Hermes profiles A to Z: five ways to use profiles, from a work/personal split to exporting your whole setup to a new machine.
• community pulse #3: how your Hermes Agent learns: the self-improvement loop and every control you have over it, from nudge intervals to the write-approval gates.
• the /journey star map: Hermes Desktop draws your agent's accumulated skills and memories as a live panel; mine is a 53-second video in the post.
• a fully local voice stack: faster-whisper for ears, Piper or KittenTTS for the voice, a local model in the middle; no audio leaves the machine.
• four ways to talk to Hermes Agent: CLI voice mode, Telegram voice notes, Discord voice channels, and how to pick a TTS provider.
• one agent, every model: my own 30-day usage data shows 8 models from 5 labs doing the work behind one agent.
• the sovereign-gpu recipe: the fourth hermes-recipes entry, a self-hosted GPU inference stack for your agent.
• drain and restore, borrow the GPU back: how my always-on box hands the GPU to benchmarks at night and takes it back for the agent in the morning.
• docs update day: six fresh docs pages in one merge, and a reminder that your Hermes Agent ships a skill that reads its own live docs.
• Hermes Agent on the OpenRouter app leaderboard: 32.5T tokens all-time and #1 daily global at capture time.

around the ecosystem

a month with two releases in it:

• Hermes Agent v0.19.0 "Quicksilver": the speed release; my highlights over the full changelog.
• Hermes Agent v0.20.0 "Herald": wake-word voice, artifacts and more; the 2-minute video changelog is in the quoted post.
• Browser Use mode: the new browser automation mode, now the default (see tip #46).
• kanban, the first official Hermes Desktop plugin: a board your agent can work.
• portable agent plugins v1: plugins as shareable packages.
• the plugin expansion wave: 29 PRs and 26 new hooks from 15 contributors, landed in one push.
• HUD mode for Hermes Desktop: a small always-on-top Hermes Agent over whatever you are working in, cmd+shift+H away.
• the in-app browser: web pages open straight inside Hermes Desktop.
• inline prompt editing in the TUI: edit any earlier line of your prompt without retyping it.
• streaming TTS: the voice answers as it thinks, and you can interrupt it mid-sentence.
• profile export and import: portable, shareable profiles with credentials stripped.
• the egress credential firewall: your API keys are off the table in Hermes Agent; tool traffic goes through a proxy that holds the secrets.
• tiered tool disclosure: big MCP servers stop being a token budget decision; tools surface as needed.
• the office document skills bundle: docx, xlsx, pdf and pptx handling built in.
• the har-derived-api-client skill: your agent watches a browser session once, then skips the browser and calls the API directly.
• Hermes inside Buzz Desktop: Block's Buzz Desktop v0.5.0 runs your Hermes Agent over ACP, same config, skills and memory.
• Hermes as a first-class Centaur harness: Paradigm's Centaur runs your agent on infrastructure you own.
• Actual Computer as a first-class provider: hosted relay or a local daemon, with a bundled setup skill.
• FLUX 3 preview: your Hermes Agent can make films now, and there is a short film contest to prove it.
• the Unreal Engine MCP companion skill: build and dress a level, place actors, frame cameras and render, from a prompt.

community corner

the part of the month I enjoy most:

• Hermes Atlas by @KSimback: the community ecosystem map, now with a use-case generator.
• the channels-intake loop, closed: a community answer to one of my intake questions ended up credited in a merged PR.
• Hermes Skins Pack v2.0: 100 CLI skins by @BChopLXXXII; I load-tested two before sharing, they drop straight in.
• the med-reminder story: a cron job that pings Telegram every 15 minutes until confirmed took one user from missed medications every week to zero in months.
• Hermes Agent on a Pi 4 with 2GB of RAM: Telegram in front, remote models behind, running fine on the smallest box anyone has reported yet.
• a full outbound GTM agent, written up: a proper long-form answer to "is Hermes useful for business".
• a superb HUD mode walkthrough video: @tonbistudio covered the feature the day after it shipped.
• a channel worth your subscription: practical AI content on agents, RAG and fine-tuning.
• Composio's harness benchmark: Kimi K3 on 28 identical tasks across three harnesses; Hermes was the fastest at a 179s median.
• GMI Cloud's harness benchmark: three models through Hermes Agent and OpenCode; Hermes never scored lower.
• LFM2.5-2.6B was trained inside the Hermes Agent harness: Liquid AI ran their agentic RL through Hermes as an unmodified black box.

local AI beyond Hermes

the 5090 had its busiest month yet: roughly one benchmark treatment a day since the 9th, every number published to the leaderboard:

• the GGUF quant sweep: "run the biggest quant that fits your card", tested on Qwen3.6-27B from Q8 to Q3. Q6_K matches Q8 exactly; Q4_K_M is the sweet spot.
• NVFP4, and this time the pitch holds: a real NVFP4 GGUF lands on the quality frontier of my K-quant ladder.
• BTL-3-Compact, the honest miss: the new sub-2.5-bit 27B trails a plain Q3_K_M on quality, size and speed.
• 15 new sm_120 field guide entries: a month of lab logs, from sglang serving hybrid-GDN AWQ broken to a CUDA build silently running on CPU.
• Ling-3.0-flash before upstream could load it: a 127B MoE on one card via a llama.cpp PR build, 46 tok/s at 3.8GB of VRAM.
• mapping --n-cpu-moe: the offload optimum is about 4GB below the VRAM ceiling, not at it.
• training your own speculative-decode drafter: an EAGLE-3 head for Gemma-3-27B, trained from scratch in four nights on the same card that serves it.
• Muse Glimmer on day 0: Meta's return to open weights, a dense 30B VLM at 85.7 tok/s fully in VRAM.
• a 2-bit quant that trades 0.9 points for 6% more speed: EschaLabs' Escha-W2 of Qwen3.6-35B-A3B, through their own runtime.
• the fastest model on my board: Nemotron 3.5 Lightning at 377 tok/s, ten quality points under the Qwen it outruns.
• all seven Lightning quants, measured: the quant tax is nearly zero; IQ2_M is the download.
• NVIDIA's own drafters make Lightning slower: both bundled speculative-decode heads bench negative, because the target is already too fast to draft for.
• and the upstream round-up: six PRs merged into SpecForge and ODS over the summer.

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

that clears the backlog. if one of these tips saved you an evening, forward this to the friend still running their browser tools the slow way.