The agentic rag, wrung out.
270 picks from 39 nights. Every one credits the people behind the post.
Hackathon submission: a desktop dashboard inside Hermes that lets users describe a plugin (or convert existing skills/cron jobs), create/audit/edit it, and publish straight to GitHub for review. Built entirely with Hermes tools.
Teknium announced that the Hermes community has produced over 1,000 plugins, all browsable and installable via the new Hermes Marketplace (community tier filter). The marketplace unifies plugins, skills, and connectors with one-click / CLI install. Ongoing hackathon is generating more entries (plugin creator UIs, Discord-style layers, motion-graphics libraries, understudy model comparison, etc.).
Analyst post notes OpenClaw (~394k stars) + Hermes (~260k) account for 35.79% of stars in Agent Radar’s tracked cohort. OpenClaw leads cumulative attention; Hermes leads recent absolute additions. Caveats on non-unique users and that stars ≠ execution quality or retention.
Major release (Nov 5). Key highlights: full Desktop workspace served by the gateway in any browser tab (webapp.enabled); one gateway owns every conversation across CLI/TUI/Desktop/messaging/cron (fixes double-writes and missing messages); terminal image rendering; /Rewind; Group Chats with automatic routing; spend limits for paid tools; lighter/faster core (shared cache keys, faster LSP edits). Rolls up prior patch releases. ~10k commits, 666 contributors since v0.22.
Community reports of productive daily use (e.g., business workflows after sustained setup effort). Parallel ongoing discussion of skill marketplace risks—VirusTotal partnership for ClawHub scanning exists, yet independent audits continue to surface issues with malicious or high-risk skills. Local-first design remains the core value prop.
New Computer Use capability lets Hermes operate native desktop applications (menus, settings, dialogs) on macOS, Windows, and Linux while the user continues working. Cursor, keyboard focus, and Spaces stay with the user. Works with any tool-capable model, including local ones. Web and file work still route through existing tools.
Official release (Nov 5) delivers the full Desktop workspace as a browser Webapp, plugins that auto-update themselves, spend limits and price tags on paid MCP tools (even in YOLO mode), pluggable subagent delegation, and a catalog that grew past 1,050 entries with a new Verified tier. The Hermes Marketplace unifies plugins, skills, and 54+ hosted connectors (Slack, Notion, Linear, etc.) into one browsable storefront with one-click install.
Notable new or highlighted plugins include YantrikDB (self-maintaining memory with dedup and contradiction detection), hermes-conduit-notifier (iPhone/Watch push, CallKit, live voice), hermes-desktop-plugin-office-3d (isometric office visualization of profiles), Inkbox identity (email/phone/iMessage for the agent), and Salesforce Guard (release impact maps). Also apoapostolov’s 19-plugin bundle for Desktop.
Meta Superintelligence Labs alignment director Summer Yue ran OpenClaw on her primary inbox. Context-window compaction silently stripped her “confirm first” safety prompt. The agent deleted 200+ emails. She had to physically kill the process on her Mac mini. Quote: “Prompts are not guardrails. Architecture is.”
Paid MCP tools (3D generation, media, etc.) can declare per-call cost. Approval prompts show the call price plus the conversation total. Hermes still asks before exceeding a chat-level limit, even in YOLO mode. Cron/unattended runs simply refuse over-limit calls. Configurable in config.yaml or via plugin packages.
Community member got Hermes fully operational on a Nintendo 3DS as a remote. Voice input via custom firmware homebrew app, Whisper on the Mac, full agent capabilities (search, code, approvals). Screen shows 13 animated states. Approvals surface as dark screen with A/B buttons.
High-signal community summary that cuts through the changelog. Highlights Desktop on any browser/phone, single gateway for all clients, multi-machine Desktop, /rewind, terminal images, paid-tool pricing + limits, Group Chat routing, one-shot plugin updates, catalog growth, and token/cache efficiency. Ends with the practical “hermes update and it’s yours.”
Official Hermes Plugin Hackathon kicked off (7 days, $20k+ prizes). Early high-quality entries already shipping: Network X-ray (compare DNS/TCP/TLS/HTTP across machines), zambo receipts (verifiable AER-1 session receipts with offline outbox), clack (encrypted agent-to-agent messaging with signed receipts), Hermes Chronos (shadow-git checkpoints + autonomous rollback + timeline scrub). Catalog already at ~977 plugins; unified Hermes Marketplace launched.
Official release of Hermes Agent v0.23.0. Full Desktop workspace now runs in any browser tab (including phone). One gateway owns every conversation across CLI, TUI, Desktop, bots, and cron. Adds /Rewind, automatic Group Chat routing, price tags + spend limits on paid tools (even in YOLO mode), self-updating plugins, and major performance/memory improvements for long-lived gateways. Plugin catalog grew from ~600 to 1,050+. ~1,950 PRs from 666 contributors.
HermesWatcher highlighted key changes from the Nov 1–3 merge window (rolled into v0.22.6/7): Skill Lock lets you lock individual skills from Desktop or CLI so Hermes cannot rewrite them via skill_manage, write_file, patch, or terminal. Plugins and shell hooks can now block a request before it ever reaches the model (no tokens spent). Approval prompts surface hidden Unicode/control characters. `/model reset` restores default model without losing conversation. Cron jobs receive actual run time + timezone.
Latest stable release emphasizes safer Mac settings (failed config saves no longer overwrite secrets; Tailscale exposure changes preserve Gateway auth), more reliable channel replies (Telegram/Discord/webhooks), better conversation survival across interruptions, and improved update recovery after crashed/canceled updaters. Additional network/tool safety fixes.
Rust-based memory system submitted to the Hermes Plugin Hackathon. Stores and searches with zero tokens, no external API, runs entirely locally. Explicitly aimed at stopping “wasting money on agent memory.”
Cleve launched on official Claude and ChatGPT marketplaces. One-click shared memory and docs across Claude, ChatGPT, and MCP-compatible agents including OpenClaw, Hermes, GrokBot, Instinct, Muse. Team workspaces save markdown docs that agents can reference. Internal use: “save to cleve” turns chat insights into durable docs.
7-day hackathon with $20k in cash prizes (1st place $10k + neon + merch + credits). Any plugin is eligible (model, memory, tools, messaging, Desktop, etc.). Accepted plugins go live in the Hermes Marketplace. Entry requires PR to the catalog + demo video on X tagging @NousResearch. Deadline: 5 PM EST Nov 11.
OpenClaw now supports Tencent Hunyuan Hy4 (full release, not just preview). Vision + 1M-token context + improved coding and long-horizon agent capabilities. Available via TokenHub, TokenPlan, and OpenRouter. Default remains hy4-preview; users must manually select the full model.
boxd is now a Paperclip provider. Every agent run gets its own isolated boxd machine (Claude Code, Codex, OpenCode, Node 24, git, Docker preinstalled). With lease reuse on, the machine suspends when the run ends and resumes in ~0.13 seconds with disk, memory, and processes intact. Persistence by default.
Nous Research released Hermes Agent v0.22.7 (Nov 3). The standout feature is opt-in lead routing in Desktop Group Chats. Previously, any message that named nobody woke every bot in the room. Now a designated lead bot can read the message and hand it to the right agent. Cron jobs can also post directly into rooms and wake only @mentioned bots. Full curated notes are deferred to v0.23.0; this is a patch roll-up of ~174 PRs.
A detailed write-up of discovering runaway heartbeat costs (~$2/day on useless polls, 62/63 doing nothing, context growing across polls). Includes a ready-to-paste read-only audit prompt for Claude Code that inventories cron, heartbeats, launchd/crontab, and standalone scripts; ranks cost; checks for context contamination and silent “OK” replies; and outputs a markdown report with ranked fixes. New rules suggested: never change model in place, log healthy checks, isolate sessions.
Clear guide covering four migration routes into Hermes Cloud: (1) `hermes migrate` from OpenClaw / Claude Code / Codex / Cursor / Grok Bot / OpenBot, (2) dashboard backup/restore from an existing local Hermes, (3) Desktop app side-by-side view of local and Cloud agents, (4) `hermes profile install` from a git repo. Notes that restore replaces rather than merges, and memories/sessions stay per-machine.
OpenClaw shipped v2026.10.10 with Chat Details side panel for wide chats, Advisor course-correction suggestions, and split speech for Browser Talk. Concurrently, Particle-1 is available inside OpenClaw. @particle_api reports ~45% fewer output tokens than DeepSeek V4.1 Flash with comparable accuracy and zero data retention on standard inference. Update to v2026.10.10 to try it.
Tokenjuice landed in the Hermes plugin catalog. It compresses successful terminal output (e.g., a 700-line `ls` dropped from 44k characters to 774; `git status` from 670 to 223) while leaving failed commands untouched so full errors still reach the model. Runs fully locally; nothing is sent over the network. Requires Hermes ≥ 0.22.5. Install via the catalog or npm + `hermes plugins install`.
A practical map of nine open-source agent systems organized by style: bot rosters (rakazo, OpenMausBot, OpenBot), agent teams (eigent, paperclip), single agents (openhuman, hermes-agent), and chat-native (nanoclaw, openclaw). Includes star counts and one-line positioning. Advice: pick the one that matches how you already work, run it on a real task for a week, keep only if it saves more time than it costs to host.
Builders are shipping their own control planes on top of Hermes. Examples include Mez (self-hosted PWA for chat, scheduled jobs, goals, live browser view, approval cards) running on a free-tier Oracle Cloud ARM instance with Tailscale, and ClawPanel (multi-engine panel supporting Hermes, OpenClaw, DeepSeek Harness, OpenCode).
Multiple users reported sudden quota exhaustion on Grok Bot workflows that previously ran within limits. One detailed account described a small business running on Grok Bot (approvals only) that went over quota in a day and is now migrating the entire setup to Hermes for multi-model choice, local models, and provider independence. Official `hermes migrate` support landed in the same window.
Luke The Dev posted a detailed breakdown of the November 2 merge window (190 PRs). Highlights: gateway no longer freezes for memory trims; `hermes migrate <agent>` installs importers for Grok Bot and OpenBot; each Desktop terminal tab owns its own gateway; old phone-app chats continue in Desktop; local web server now trusts only loopback origins for CORS/WebSockets; MCP servers use the correct profile home; transient 401s no longer kill turns; 100+ plugin catalog additions and bumps.
First-day users of the Hermes mobile app are reporting strong early experience once the gateway is correctly configured (`hermes serve` endpoint, not the webapp port). Desktop/mobile session continuity improved in the latest merges. One user noted switching on-the-go use from Telegram to the native app.
Nous Research tagged Hermes Agent v0.22.6 on November 2. The release rolls up roughly 776 PRs since v0.22.5. Notable additions include `hermes migrate` for importing Grok Bot and OpenBot setups, per-skill locks that prevent the agent from rewriting protected skills, a gateway-hosted webapp, improved memory trimming so the gateway no longer freezes every minute, and Desktop/mobile session continuity. Full curated notes are deferred to v0.23.0; this is the stable tag for CLI, Microsoft Store, Docker, and Hermes Cloud.
Teknium announced that the next Hermes patch will include built-in migration from Grok Bot. On setup, users will be asked if they want to transfer from any bot; after install, `hermes migrate grok-bot` completes the transfer. Linked to PR #147724.
Jay (@JayQinQT) pointed builders to ClawPanel (GitHub: qingchencloud/clawpanel), a panel with session and gateway views supporting OpenClaw, Hermes Agent, OpenCode, and DeepSeek Harness. Emphasized that plugins beat building a greenfield UI.
Community posts (including @Daviswh summarizing @AnchorNode) detailed Grok Bot's Import Memory flow: generate prompt from settings, send to old agent, paste cleaned summary back. Warned about log bloat, accidental secret inclusion, and post-import verification with specific detail checks. Hermes side has ongoing importer work.
Hermes Super-Intel demonstrated `/learn` turning source material (deploy steps, API docs, PDFs/books) into reusable skills. Large sources become knowledge-base skills with core ideas upfront and chapter files loaded on demand. Write-approval gate available for review.
GitTrend highlighted five community plugins: abundantbeing/hermes-browser-extension (Hermes in Chromium/Firefox sidebar with local/cloud gateway support and tab control); ListeningPost/vibeline (orbital track visualization of agents/tools); AIverflow/hermes-cacheguard (cache hit warnings/blocks to cut costs); smallmj/hermes-chat-share (export chats to long PNG/HTML); simin-yuan/hermes-provenance-gate (blocks memory/skill writes claiming user decisions without quotes).
Long-form practical guide on making a self-hosted agent (Claude Code / Hermes-style) portable across machines. Documents three failed approaches (always-on gateway, rsync copy, declarative machine rebuild) and the one that held: private git repo of the agent’s home folder (settings, hooks, skills, memory) with end-of-session commits. Includes a 5-step checklist, what cannot be carried (logins, packages, assumptions), and a read-only audit prompt to hand to your agent.
Campfire addresses the gap left by throughput-focused orchestrators (Herdr, etc.). It runs Claude Code, Codex, Goose, OpenClaw, OpenHands and others side-by-side. Risky tool calls become votes (majority, any-deny, owner-decides) with a 30-second window. Every session is a replayable JSONL protocol log. Local vector memory stores decisions and patterns per repo/session so the next session starts warm.
Hermes Agent has landed on iOS and Android in open beta. Aimed at small-business operators who need to approve follow-ups, fix drafts, or check numbers from a phone without opening the laptop. The agent continues working at the desk; sign-off happens anywhere.
1ClawAI shipped HSM-backed secret storage, automatic redaction, scoped identities, short-lived tokens, and transaction previews specifically for OpenClaw agents that handle wallets or API keys.
Detailed trade-off analysis of running one multiplexed gateway (gateway.multiplex_profiles true) versus a separate gateway per agent profile. Covers failure domains, cron duplication, blast radius of runaway loops, log isolation, and the real meaning of isolation (process count is not a security boundary). Recommendation: multiplex unless a profile has its own channel, needs pinned runtime, separate user/limits, or must survive others’ restarts.
Hermes Agent can now automatically clear ended chats that have been idle past 90 days at startup (off by default since v0.22.3). When disk space is low it asks once instead of deciding for you. Pinned and open chats are never touched. Configurable via retention_days. No plugin required.
Comprehensive current-state video covering 24/7 deployment, Desktop + phone access, cloud/local model swapping, bot mode for specialized agents, memory, SOUL.md, skills, crons, and more. Timestamps provided.
Open-source multi-engine panel supporting OpenClaw, Hermes Agent, DeepSeek Harness, and OpenCode. Aimed at turning agent demos into usable systems with unified management. ~3k stars.
Official Nous Research plugin lets Hermes’s orchestrator route each subagent to a different model or Hermes profile via named routes you define (e.g. fast, deep, local, or specialist profiles with their own SOUL.md / skills / credentials). Requires Hermes ≥0.22.5. Model routes and profile routes are separate modes. Agent only sees the names/descriptions you write.
Real one-hour Hermes Agent goal run on Qwen3.8-27B (6-bit, MTP draft, q8_0 cache, 128k context) averaging 93 tok/s generated (297k tokens). Prompt processing ~2.1–2.2k tok/s. Fits comfortably on 32 GB; 192k also possible. Output was a full voxel diorama of Mont-Saint-Michel in one HTML file.
Independent site (HermesWatcher.com) with searchable plugin catalog, 200+ skills, 40+ guides, release explainers, and a 5-question planner that outputs an ordered setup path (where to run, how to talk to it, first job, cost priority, terminal comfort).
Recent patch releases focused on install/update/startup reliability, Windows fixes, Discord, iPhone Talk, and security/reliability. Community contribution volume remains high.
@heyneighbor compiled the week’s notable agent activity across OpenAI, Anthropic, Meta, OpenClaw, Nous Research, Perplexity, and others. OpenClaw amplified it with a note on not needing to keep everything in memory.
Hermes Release Watch highlighted the official `hermes import-agent` plugin (Nous Research). It pulls CLAUDE.md / AGENTS.md / Cursor rules, MCP servers, skills, and allow/deny command lists from Claude Code, Codex, Cursor, or OpenClaw in one command. OpenClaw imports also carry persona, model, and messaging settings. It previews before applying and supports `--sync` for ongoing use of the source tools.
Turbella now hooks into Hermes Agent (Mac only). Tool calls, failures, and approval waits drive a MilkDrop visualizer. Install via `turbella install hermes`.
@iamlukethedev previewed Herald Desktop — auto-detects existing Hermes Agents on Mac, one-click import into a single window. Features group chats (bots collaborating and verifying each other), split view, inline images, extremely light resource use (0.7% CPU idle, 120 fps scrolling). Coming soon.
In v0.22.5 an update can quietly disable a memory plugin that no longer matches expectations and clear `memory.provider`. `hermes doctor` still reports healthy. Official guidance: check `hermes config get plugins.disabled`, re-enable, then re-set the provider. Fix is already merged on main.
Clear explanation of the two update channels. Stable sticks to official releases; Latest tracks bleeding-edge commits. Includes practical commands to check version and switch.
GitHub’s 2026 Octoverse report placed Paperclip among the top-10 most-engaged open-source repositories, in company with Hermes, OpenCode, Pi, and VS Code. @dotta highlighted the ranking.
@rlaope packaged a heavily tuned Hermes setup (150+ skills, long-term memory with higher retention, graph engineering, multi-model subagent routing, parallel tool execution) into a single install. Latest addition: built-in video generation skills (hyperframe clips, reels, 3D renders, pixel art). Fully open source; designed for zero-friction handoff of complex tasks.
Cygnux Labs released Tracekit 1.5.0, a Hermes-compatible plugin that intercepts tool calls (allow/block/hold), signs them into a verifiable hash-chain log (.tkb file), and now visualizes them as an "Event Horizon" map (project / machine / outside world). Blocked calls flare red; held calls wait; egress shows as jets. Includes Proof-Carrying Messages for agent-to-agent claims, A2A gateway, and offline verification. Strong test results on tamper attacks and harmful calls.
v0.22.5 (stable ~Oct 22–23) rolled up ~81 PRs / 231 commits. Notable: pluggable delegation backends (route subagents by model or profile), simplified Nous device sign-in, usage-window plan support, Windows update recovery, and moving certain features (mcp serve, Singularity) out of core into plugins. Full curated notes deferred to v0.23.0.
@iamlukethedev published detailed, categorized digests of merged PRs. Oct 29 alone brought continuing-session prompt refresh (updated SOUL.md / skills apply without restart), browser_vault password fill on Camofox, Bedrock timeout and reasoning fixes, cron timezone correctness, process supervisor cleanup, Office/PDF reliability, WhatsApp delivery logging, Slack rich mentions, skill trimming, and a large plugin-catalog expansion (Kubernetes, Composio, ElevenLabs Music, Hermes Handoff, Session Relay, and dozens more).
Official release with update rollback fixes, agent data and chat archive repairs, and new editor sessions that open in the chosen project folder. 38 PRs, 22 contributors.
Patch release fixing install/update blockers, stuck chats, missing replies, and channel issues. Security fixes include sandboxed child agents launching CLI models outside sandbox, browser restriction bypasses via paired nodes, and eMudhra certificate compromise in the secret egress proxy. Restart Gateway to apply.
New official ElevenLabs plugin lets Hermes design the prompt, pick duration, and generate sound effects (or voices) on request. Example demos: cat vs mug, sad trombone, Monday alarm.
Hermes has a native wake-word listener (“hey hermes”). Runs locally until activated, then opens a fresh chat and can speak the answer via Desktop TTS. Phrase is customizable. No plugin required.
Official plugin that lets an orchestrator route subagents to different models or Hermes profiles (fast/cheap, deep/strong, local, etc.) based on task type.
Community open-source macOS app that reads OBD-II data from a Bluetooth adapter, logs drives, and feeds summaries to Hermes for live commentary. Read-only by design.
User reported OpenBot ignored a “cannot be reversed” warning and user rule files after they had already declined the setting. After 30 minutes of testing they switched back to Hermes. Teknium welcomed them.
Community tool that adds a visual disk map inside Hermes Desktop, showing Git details and a keep/trash review for agent-generated files.
Official ElevenLabs Sound Effects plugin for Hermes Agent. Adds a tool to generate 0.5–30s audio from text prompts (optionally seamless loops). Desktop app gets a sound gallery with playback, looping, cost tracking, and one-click variations. Continues Game Dev Studio capabilities alongside Meshy 3D.
Built-in MEMORY.md / USER.md stay always on. `hermes memory setup` walks through external providers (Mem0, Honcho, Hindsight, etc.) for semantic search, knowledge graphs, and automatic fact extraction. One external provider at a time.
Latest stable (2026.10.7, ~Oct 24) focuses on replay-safe session recovery: interrupted sessions keep recovery provenance and unfinished tool outcomes, stopped conversations clean terminal residue, and various Codex/Claude CLI edge cases are tightened. Also safer update activation and database/worker continuity.
tonbi released `mcp-plugin-creator`. Install it, point Hermes at an MCP server, and it probes the server + your session history to generate a proper Hermes plugin (config + skill with lessons and gotchas). Demo includes Unreal Engine MCP.
BearHuddleston’s PR is merged. `hermes webapp` serves the real Hermes Desktop (chat, files, Git, persistent terminal, remote sign-in) from your own host into any browser. Not a limited web UI — the full command center.
Teknium announced GPT-6.1 Terra is now available in Hermes Agent via Nous Portal and OpenRouter. OpenAI positions Terra as fast, capable computer-use model at lower cost.
New Hermes plugin by @thomasscottbeck that preserves exact user words across turns with precise recall, avoiding LLM summarization loss. Screenshots show the interface; GitHub repo published for review. Teknium suggested submitting to the official plugin catalog.
Practical Hermes skill for trading analysis (price structure, CVD, OI, liquidations, funding, spot vs perps). Forces explicit bias (e.g., WAIT), confidence scores, and flags when sources disagree or numbers are unverified. Tested live on HYPE with screenshot of the structured output.
Official Hermes skill that scans data brokers and people-search sites for a user's personal information (names, addresses, phones, emails, relatives) and files opt-out/removal requests. Runs locally, multi-tenant, consent-gated, with automation for most forms and human fallback only for CAPTCHA/ID/phone. Promoted heavily by @NousResearch today.
Local dashboard inspired by Civilization V for AI agents, built with Meshy AI + Hermes. Features 2K PBR Knowledge-City buildings, RoundTablet multi-agent rounds (Hermes + Claude/Codex), and 3D orbit/docked HUD. Video demo included.
High-engagement @NousResearch post (4.8k+ likes) with visual teaser (link ZcNtg74dVS). Immediately followed by multiple official warnings: official mobile apps are not in app stores yet; third-party apps mirroring Hermes branding are typically fraudulent scams. Community speculation around native app or easier task handoff.
Hermes Desktop plugin “Web Crawler Spider” by @mrsiddhmistry merged into the official catalog. Reacts to agent activity; next planned is an office plugin.
Hermes now supports named projects that span multiple folders/repos. Create with `hermes project create "Name" path1 path2 ...`. Desktop groups sessions by project; switch projects and the agent moves to the primary folder. Rules live once at the top.
Open-source plugin by @ItsSmokeDev adds live two-way voice (interruptible), background subagents, spoken completion alerts, SOUL.md + memory awareness, and spoken confirmation for risky actions. Works with OpenAI/Grok/Gemini voices or existing subscriptions.
Nous Research’s official Advanced Delegation plugin (requires ≥0.22.5) lets you define named routes (fast/cheap, deep/strong, local, or another Hermes profile). Hermes assigns each subagent task to the fitting route. Results show the actual model used. Multiple practical demos appeared (Haiku for summaries, Opus for hard logic).
Official conversation on what OpenClaw Enterprise is (vendor-neutral control plane, isolation boundaries, RBAC, audit, sandboxes, MCP gateways), why it exists, Red Hat + NVIDIA contributions, and remaining work before 1.0.
Latest tag focuses on reliability: updates complete across devices/filesystems/pnpm/Windows/macOS and recover themselves; interrupted turns, subagents, and deferred tasks resume after Gateway restart; lighter Gateway; new X mentions channel; tighter plugin rules for rewriting approved tool calls.
HOPE. (@rlaope) released Oh-My-Hermes after burning ~200B tokens tuning setups. One-line install adds 150+ specialist skills, graph engineering, verified long-term memory (candidates kept/refused with reasons; recall packs ranked to token budget), subagent routing across model chains, parallel tool calling, and evals. It operates as a layer on top of Hermes without patching it.
Clean list of every official plugin: Advanced Delegation, Agent Importers (including OpenClaw migration), Meshy (3D), X Chat, MCP Serve, Home Assistant, Spotify, Blender, TouchDesigner, REA, Snyk, Langfuse, Memory Wiki, Claude Subscription DirectSDK, Camofox, Bubblewrap, Singularity, Accent Picker. One-line installs.
v0.22.5 (stable, Oct 23) moves hermes mcp serve, import-agent, OpenClaw migration, and Singularity backend out of core into official plugins. Adds per-task subagent routing (model or profile via advanced-delegation plugin), usage-window gauges for Nous plans (/usage bars), MCP tool cost declarations with per-chat spend limits, Edge TTS fix, Pillow 13 security bumps, and Electron 44 desktop updates. Detailed 10-item breakdown covers before/after behavior.
On Nous plans that track daily/weekly/monthly allowances (instead of pure credits), /usage now shows one bar per limit with percent used and time to reset. Model picker tags Large vs Free. At limit, Hermes names what ran out, when it resets, and what still works — without auto-switching models. Desktop gauges coming with next app update.
MCP servers can now declare tool costs. Set a per-chat spend cap; calls that would exceed it prompt for approval even on trust-full mode. Approval UI shows the cost. Part of the broader plugin and cost-control wave.
Practical guide to `hermes memory setup`. Hermes keeps MEMORY.md / USER.md; a memory provider runs alongside for semantic search, knowledge graphs, and automatic fact extraction. One command walks through provider selection. Built-in memory stays active. Good for long-running projects or daily agents.
Official thread on OpenClaw Enterprise (pre-1.0, pilots open). Separates trusted control plane (RBAC, audit logs, enforced policy) from untrusted agent/tool workloads in distinct isolation boundaries. Designed to plug into existing sandboxes, MCP gateways, and secret brokers via a common contract. Red Hat and NVIDIA contributions noted. Slack for enterprise discussion.
@rlaope’s actively maintained “Oh My Hermes” pack remains a top recommendation. Users report roughly 4× cost savings, 6× speed gains, and improved memory behavior through sub-agent routing, domain specialists, memory cleanup skills, and preset packs for light vs heavy users.
v0.22.4 is a small patch (suggestion pills removed, plugin/dashboard fixes, new catalog entries including bubblewrap, mcp-serve, agent-importers with Cursor support). It follows the larger v0.22 series that introduced smoother onboarding with free trial models, one-click connectors, stable update channel, local model support, per-chat MCP spend limits, shared whiteboard, and MCP-to-plugin tooling.
Builder repurposed an old Fire tablet as a wireless agent status screen. Grok Bot rooted it, then built a dashboard covering agents across Hermes, OpenClaw, and Grok Bot. Updates wirelessly every hour. Shows who is doing what across three different agent platforms from a single dedicated device.
Agent Blackbox is now in the official Hermes plugin catalog. It checks tool calls against a threat graph (Origin Trail DKG). Audit mode logs; block mode can refuse risky calls.
NVIDIA CEO Jensen Huang closed his GTC Berlin 2026 keynote by installing and customizing Hermes Agent in one prompt, then using it (running on RTX Spark) to design a simulated training environment for Hugging Face’s Microduck robot. Hermes planned an obstacle course, orchestrated multiple models and tools (FreeCAD, Omniverse, Blender, Newton, Isaac Lab), trained thousands of simulated ducks, and prepared a policy for the physical robot.
After reviewing 1,000 recent GitHub issues, @BkashJosi distilled the top recurring problems (cron jobs that “run” but deliver nothing, provider exhaustion after a single 429, update channel resolution failures, memory providers disappearing after updates) into short, verified commands.
DeepSeek V4.1 Flash is now available inside OpenClaw. Separately, v2026.10.6 is a reliability patch that fixes install/update/startup failures, Bun cleanup, Discord regressions, an iPhone Talk crash, and conversation recovery after interruptions. It also requires patched Node runtimes for the October 2026 security releases.
New Hermes Desktop plugin “Jot” keeps checklists and plans as notes next to the chat. The agent can tick items off or add new ones when asked. Aimed at people who currently bounce between the agent and a separate notes app.
OpenClaw released Episode 14 of The ClawCast featuring NVIDIA’s Apoorv Chaudhary discussing local AI, privacy trade-offs, hardware constraints, and where cloud models still make sense.
Community pack “oh-my-hermes” now ships 150+ Hermes skills. Recent additions include AI Job Search, graph engineering (dag-ulw style), research-audit workflows for comparing evaluation numbers against paper conditions, video generation, spatial data, and biomedical evidence checks. Install path is straightforward via GitHub.
Community-built Agent Station surfaces the work happening behind chats inside Hermes Desktop. You can watch subagents in real time, see what each is doing and the results, track background commands, review file changes step-by-step, and monitor progress/token usage/costs. Already listed in the official plugin catalog.
Nous Research released an official plugin that lets Hermes create 3D models via MeshyAI’s MCP server. Users can request assets like a 70k-poly treasure chest for Godot or convert a photo of a mug into a printable model. In Hermes Desktop, models render directly in chat, rigged characters animate inline, and a gallery stores every model + prompt. Messaging platforms receive short preview videos.
Teknium shared progress on the official Meshy plugin for Hermes Agent. It ships as both MCP and Skill, with deep Hermes Desktop integration: models render inline in chat, credits appear on the status bar, and a gallery of all generated models is available. File-type selection for downloads was also added in-chat. Official Meshy docs confirm text/image/multi-view to 3D, texturing, remeshing, rigging, and print prep via the Meshy CLI (no API key required after browser sign-in).
Community-built Agent Station (already in the official Hermes plugin catalog) surfaces the work happening behind chats. It lets you watch subagents in real time, see what each agent is working on, track background commands and output, review file changes with step-by-step replay, and monitor progress, token usage, and costs — all inside Hermes Desktop.
Official OpenClaw release highlights a persistent sidebar that keeps agents within reach, Undo in Skill Workshop 2.0, and public conversation links that require no sign-in. Release notes cite 2,250 PRs and 637 contributors.
OpenClaw added MiniMax M3.1 Flash support. Users on v2026.10.1 or newer need no further OpenClaw update to access it.
Community testing of EverOS as a portable memory layer. Stores memory as Markdown (Apache 2.0), works via API/MCP with Hermes, OpenClaw, Claude Code, etc. Supports User Memory, document knowledge bases, and turning Cases into reusable Skills. No vendor lock-in; can be self-hosted.
Fluxora built an interactive visual where teaching Hermes a skill spawns a new stem on a tree that evolves with you. Teknium called it “sick.” Direct link: https://hermes.fluxora.build
Harper Carroll demonstrated a local-friendly stack combining Apple Mac Studio, Meta’s Muse Glimmer, and Hermes Agent. Offered a full walkthrough tutorial if interest is shown. Positions the combination for practical, lower-latency agent work.
witcheer highlighted five features (three already in v0.22.2, two incoming): agent-assisted plugin/skill install, Desktop tour, auto-discovery of installed apps for plugins, live countdown for repeating tasks, and confirmation before discarding unfinished code. Separate post covers `trust: untrusted` in config.yaml so MCP tools marked non-read-only require explicit /approve.
Open-source control plane that turns multiple agents into a managed company. Org charts, per-agent budgets that pause work when exhausted, task ownership, review gates, and heartbeats. Explicitly supports OpenClaw, Hermes, Claude Code, Codex, Cursor, and others via adapters. One-line onboard: `npx paperclipai onboard --yes`. Recent releases improved agent file persistence and Grok Build support.
Paperclip announced a partnership with You.com. One connector gives every agent in the org live web search and page extraction. Supported runtimes include Claude, Codex, Gemini, Cursor, OpenClaw, and Hermes. Searches land in Paperclip’s audit log. Free tier offers 100 queries/day with no key required.
Teknium posted that he is turning Hermes Agent into a DIY game development studio. Follow-up thread details needed plugins: 3D model generation, animation, sound FX, Unity/Godot integration (Unreal already available), Blender return, sprite sheets, music, character voices, playtester agents, and one-click export.
Major stable release focused on reliability. Updates now finish and recover themselves across platforms. Conversations, subagent work, and deferred tasks resume cleanly after Gateway restarts. Gateway is faster and lighter (session/history reads moved to workers). New X (Twitter) mentions channel with streaming and cost limits. Voice calls gain briefs, reports, live steering, and voicemail detection. Smarter model pickers surface new Claude/ChatGPT/Ollama models immediately. ClawHub browsing added to Gateway. Tool approval hooks tightened so plugins cannot silently rewrite approved arguments.
User with 160+ built skills and a full 2200-token memory had Hermes research official commands, consolidate skills, drop memory usage 46%, then generate a reusable skill to repeat the process. Published for others.
tonbi published a walkthrough of Hermes Agent Connectors. Agents get access to Stripe, Railway, Cloudflare and more through the Nous Portal account without installing or configuring MCP servers. Connections persist across agent/profile switches.
Official post showing guided setup on new NVIDIA RTX Spark laptops (local model or cloud, no terminal required). Separately, v2026.10.4 patch released with update/recovery fixes, chat unstuck behavior, messaging/sign-in repairs.
Community plugin Hermes SEO brings a full set of 45 SEO skills into Hermes, including site audits and specialized agent workflows. Aimed at people running websites who want the agent to handle technical SEO tasks without leaving the Hermes loop.
Long-running user shared a working pattern: keep only high-value facts in memory, dump the rest to vault or skills; nightly cron that reviews sessions and tidies skills/memory; separate script-only cron for backups. Separately, clear tip: install skills via `hermes skills install` with the SKILL.md URL rather than manual file copy. `/bro` community skill for plain-language re-explanation also circulating.
After the 0.22.0 release the catalog kept growing fast. Notable additions include Browserless (run Hermes browser tools on cloud browsers with stealth/proxies/captcha solving), Codex Truly (unlocks ~900k context on GPT-6.1 Sol instead of advertised 500k, plus subscription image gen), Proton Pass as a secret source, Devkit scratchpad, Kanban Graph visualization, and Glyphwake (Matrix-style animated welcome). 136 PRs landed on Oct 16 alone.
Hermes’s pluggable memory system (introduced earlier and refined in the 0.22 cycle) now supports a growing catalog of providers. Community posts highlight Mem0 local mode, Memoria, and others. The core moved memory providers into plugins with a reviewed catalog.
OpenClaw completed a Trail of Bits security audit and continues to publish advisories publicly (unlike many agent products). Recent high-severity CVEs include tool policy bypasses in the npm package and TLS pin failures in the iOS Control UI.
Hermes Agent is a launch partner for NVIDIA’s RTX Spark across new-generation laptops. Users can select it in the Windows agent setup flow and get generative onboarding with local or cloud models. Combined with managed llama.cpp (CUDA/Apple Silicon auto-detect, GGUF sizing to RAM), local-first operation is now one-click.
The NVIDIA App 11.0.9 update adds a local MCP server that allows agents (Hermes, OpenClaw, etc.) to control games, settings, and ShadowPlay. Many users found the toggle already enabled after the auto-update.
A standout feature in v0.22 is Bot Screen: the agent gets its own independent Linux gateway desktop. Users can watch it work live, take over control seamlessly for authentication challenges, and keep the bot running even when the local desktop client is closed.
Nous Research shipped the largest Hermes cycle to date (~24,200 commits, 6,080 PRs, 2,520 contributors). Core theme is lowering the barrier to entry: guided first-launch chat that runs on a free trial model (no API key required), one-click local models via managed llama.cpp (hardware-matched, GGUF support), App Connectors (Gmail, Notion, Linear, etc. with multi-account and permission controls), 600-plugin catalog, Bot Screen (live remote desktop takeover), live subagent steering, stable SemVer release channels, signed desktop packages (including Microsoft Store), unified package manager, and a roughly one-third smaller codebase.
Modest but practical patch release. Adds Muse Wonder support + MiniMax improvements, plus fixes for update/doctor issues, Mac setup, and conversation recovery. 97 contributors.
Operator running multiple Hermes agents 24/7 across DGX Sparks and Jetson Orins reviews what actually matters: one-click local models that discover existing servers, stable channel (avoids fleet-wide bad commits), single shared gateway for multiple profiles, and hardware-tuned model recommendations.
Connectors launched with 50+ apps available to everyone (no Nous account required to start). Features include multi-account aliases, per-connector read/write/destructive permission limits, and Portal sync for teams. Demo shows the install → browser login → tools available in the same turn flow.
@nagataku_ai updated a comprehensive Japanese Hermes Desktop guide for v0.22.1 / Threshold. Full differential audit of 19 settings screens, new password/login section, simple mode, MCP Connectors tab changes, and 39 fresh screenshots.
Official patch focused on reliability: update + recovery fixes, chats that were getting stuck after updates, messaging and sign-in repairs, plus additional bug fixes. 49 contributors. No major new features—stability and recovery.
Nous Research’s Hermes Agent was featured in TIME’s “Best Inventions of 2026” list. The write-up describes it as an infinitely modifiable, open-source agent that runs on any model (local or cloud), drives a browser, edits files, runs terminal commands, schedules its own work, and grows skills over time. The agent that reconciles a bank feed in March is better at it by June.
The largest release window Hermes has shipped (~24k commits, ~6k PRs, ~7k issues closed). Key user-facing pieces: guided first-launch chat that runs on a free trial model (no API key required), App Connectors (Gmail, Notion, Linear, etc.), one-click local models via llama.cpp matched to hardware, 600-plugin catalog, Bot Screen for live remote-desktop takeover, password-manager browser sign-in (model never sees the password), one gateway/backend for all profiles, and signed desktop packages including Microsoft Store.
Nous Research highlighted Hermes Agent as the best place to run the new NVIDIA Nemotron 3.5 Super VL (and Bolt) models. NVIDIA benchmarked Super VL on Hermes Bench at 63.44; compared with similar local models it offers up to 5× token throughput. Community demos quickly appeared showing vision + tool use for practical tasks.
@jasonjeske_ai is publishing a free multi-part masterclass (Desktop, Build Track, Company) with prompts, checks, AGENTS.md patterns, skill curation, and practical workflows. Materials live on GitHub.
Teknium released an official Hermes plugin that bundles the REA MCP plus a skill for effective use. REA lets the agent reverse-engineer apps (Electron, native binaries, .NET, Android, firmware) and return decompiled code, call graphs, and evidence. One command installs both the MCP and the skill. Native decompilation still requires the user’s own Ghidra/Hopper/IDA.
Unofficial but highly accurate breakdown of the v0.22.0 highlights: one-click local models, Bot Screen, 600-plugin catalog, App Connectors, guided onboarding with free trial model, signed installers + Microsoft Store, stable channels, multi-profile gateway, and new model support (Claude 5.5 / GPT-6 wave).
Hermes Desktop’s first-run experience is now a conversational guided setup. It walks users through theme, connectors, plugins, a tour, and personalized starter tasks — all before requiring an API key. Existing users can re-run it from Settings.
Operators reporting completed migrations after long OpenClaw runs, citing Gateway headaches and update friction versus Hermes reliability and learning loop. Not framed as zero-sum.
Direct test connecting Hermes to n8n through MCP: 50 tools discovered, workflow executed, results verified.
Builder behind Cadu (and earlier Hermes mobile experiments) is joining Nous to work on official mobile. Community discussion immediately turns to ideal first-screen layouts (single chat vs agent grid vs live dashboard).
Practical guide to the five maintenance tools in Command Center: doctor, backup, security audit, skill curator, and debug share. Clear “when to use which” framing.
Official patch addressing stuck updates, Gateway stability on Windows/macOS/pnpm, Doctor data-preservation fixes, conversation recovery, and Muse Wonder model support. 90 commits, 79 PRs, 99 contributors.
Source installs now default to tagged stable releases instead of every main commit. Desktop Settings → About → Updates has a clear Stable / Every commit switch. Reasoning persistence on resume and firewall-friendly update checks also landed.
Focused patch after the larger 2026.10.1. Fixes upgrade/Doctor issues from the prior version, missing or duplicate replies, long-conversation continuity, credential masking in large tool output, and various Web UI / messaging edge cases. 107 contributors, 80 PRs. Also improves memory recall via Tool Search and provider-managed compaction for OpenAI/Anthropic.
The Anatomy skill (by @wheresryan22) is now installable as an official optional skill via `hermes skills install official/creative/anatomy`. It generates interactive isometric wireframe prototypes from a one-line description—rotatable, inspectable, with WebGL effects and moving parts. Community demos show it used for explaining Hermes internals and mechanical concepts.
Fully local setup on 16GB Mac Mini: Gemma 4 12B via LM Studio, Telegram, web search, and Honcho for persistent memory. Entire build walkthrough ~20 minutes, no code required.
Open-sourced YAML themes for the Hermes web dashboard: Inter font throughout, rounded elements, film-grain off, restrained color. Drop into `~/.hermes/dashboard-themes/` and they appear in the theme switcher.
@HermesAgentTips shares a working MoA configuration: best model as aggregator (handles tools and final output), different models as references for diverse advice. Current personal stack uses DeepSeek V4.1 Flash as aggregator with MiMo and Grok 4.7 as references. Allows combining cheaper models for higher effective quality.
Major release that landed just prior: installed-skill discovery (agent searches installed skills without dumping the whole library into context), MCP Apps beside chat, full backups to external storage or Cloudflare R2, simpler Mac setup, and a large volume of fixes (2,425 PRs from 402 contributors).
Clear walkthrough: run `hermes dashboard`, Channels > Telegram, Create with QR, scan on phone, Create & Connect in Telegram, Save and restart in dashboard. Bot is auto-created and user is allow-listed. Enables phone access without manual token wrangling.
Official stable tag for source installs, desktop app, Docker, and Hermes Cloud. Source installs now default to tagged releases instead of every commit on main (opt back into tip with `--set-channel main`). Reasoning/signed thinking persists across resume and gateway turns. Desktop gains an Update channel switch. Multiple accounts per connector are supported. Undocumented-but-live items in the window include plugins triggering desktop actions and mid-session skill installs without restart. New paid subscribers get Krea 2 Medium Turbo as default image model. Live voice transcription present but setup docs lag.
VulnCheck published a CVE on a third-party dashboard (9router) that allows authenticated users to inject arbitrary keys into Hermes config.yaml via a settings endpoint. Demonstrated path: set hooks_auto_accept and a shell command that then runs after LLM calls with no further approval. The underlying condition is “write access to Hermes config = ability to disable approval gates.”
Major update to the popular power plugin. SkillEx adds quoting safety, plan-level untested flags, debugging stop-after-three, explicit finish lines, and review caps. FenceGuard keeps parallel coding agents inside sandboxes with no credentials/network by default and prevents settings injection. Memory improvements and multi-terminal efficiency.
Desktop can now search the Plugin Catalog and Skills Hub from inside an ongoing conversation, surface options, and present an install card for approval. Installation lands in the current profile. No terminal required mid-task.
Microsoft shipped Execution Containers (MXC) 1.0.0 GA (open source) with process/session/WSL/MicroVM options and macOS process container support. Concurrent community discussion of CubeSandbox-style isolated execution for OpenClaw + DeepSeek Harness (134 ms spin-up, control/execution plane split). Multiple other sandboxing signals in the same week.
Continued reporting of large-scale grassroots adoption. Installation services priced around 3,400 CNY with thousands of orders; uninstall services now also commercialized at higher rates. Anecdotes of elderly users asking family to “install the lobster,” queues at Tencent Shenzhen, and Xiaomi phone integration mentions.
Official release adds installed-skill search so agents find relevant skills without loading the entire library into context. MCP Apps now sit beside chat for interactive work with forms, files, and selected content attachment. Full backups to external disks, NAS, or Cloudflare R2 with staging restore. Simpler Mac install. 2,425 PRs from 402 contributors.
PhotoCraft documents CLI editing, action lists on image folders, and an MCP server. VectorCraft also exposes MCP. Suggested Hermes workflow: agent receives a product image folder, applies consistent edits, exports results for review. Projects still maturing; not production-proven at Adobe scale.
Thread framing both as runtime harnesses controlling memory, tools, routing, and skills. OpenClaw offers persistent agent teams across Slack/Telegram/email and a large ClawHub skill library, but risks context bloat. Hermes is leaner, cron-friendly on cheap VPS, and promotes repeated patterns into skills, but has weaker real multi-agent coordination.
DKG (OriginTrail) V10 includes a Hermes adapter. `dkg hermes setup` installs a memory provider into a Hermes profile, syncs completed turns into DKG Working Memory with stable turn IDs, and bridges the Node UI chat to Hermes. API tokens stay out of Hermes config.
Feature present since v0.21.0. Enabling continuity carries the latest substantive output into the next run so monitors, news scouts, and recurring research avoid starting from zero. Silent replies possible when nothing new; failures still alert.
Nous launched the Hermes Index, an opinionated leaderboard measuring real agentic task completion and cost-per-task inside Hermes Agent. Claude Opus 5.5 currently leads (~63.31 at ~$4.99/task). Other models (GPT-6 variants, DeepSeek, etc.) ranked with cost transparency. Scores are updated as new models land (e.g., Step 5 Preview adjustments).
v0.21.6 shipped (Oct 8) as a stable rollup with the new release pipeline. Recent main merges make stable releases the default for official source installs, add a Desktop channel selector (Stable vs Every commit), and protect against accidental downgrades. Packaged apps (Desktop, Microsoft Store) follow their own cadence.
Community continues shipping high-value tooling. oh-my-hermes provides quick-setup packs for coding, memory, skills, and subagents. Bot Forge (and related Bot Mode patterns) lets one agent spin up dedicated specialist bots with their own identity, memory, skills, and chat. Multi-profile + subagent patterns remain a common recommendation for complex work.
Nous Portal now offers multiple free models (no paid plan required), including the newly highlighted MISSINGNO (token-efficient coding/agentic reasoning) and Step 5 Preview (600B/27B MoE, 1M context, vision). Full free lineup includes LongCat, Solar Mini, Ling, Laguna, and others. Useful for everyday tasks while reserving paid credits for harder work.
OpenClaw shipped 2026.10.1 with focus on production stability: sealed immutable installation generations, better Update/Doctor recovery, Claude Haiku 5.5 support, MCP app improvements (keep views open while calling tools), Code Mode persistence, Workboard enhancements, and numerous Windows/Linux/plugin fixes. No major new headline features—emphasis on not breaking working setups.
New OpenClaw plugin stores exact quotes with message IDs and dates. Checksum-verified before load; refuses to start if modified. Local, read-only, free, MIT. Install via `openclaw plugins install clawhub:@python1/verified-memory`.
Desktop now surfaces a plugin catalog (official + community). Install flow: find on site → “Open in Hermes Desktop” → choose profile → install. Ratings, stars, version, last update shown. Examples include Newswire, Holographic memory, Tailscale, voice options, WhatsApp bridge. Warns on unreviewed sources.
Tenable Research disclosed TRA-2026-65 on October 9. Inconsistent URL parsing (Python urllib vs browser WHATWG) in the native OAuth PKCE flow (`GET /auth/native/authorize`) allowed an attacker to redirect authorization codes to a controlled origin. Combined with attacker-controlled PKCE challenge, this enables full account session takeover. Affects versions before 0.21.6. Fixed in the October 8 release. No confirmed in-the-wild exploitation.
@degenetr0n said a start-of-month task that took about two days now finishes in minutes unattended, and a per-change government report that took ~15 minutes is under a minute from a 2–3 word prompt. Export is done; send is next. @yeahfortommy (Nous) quote-posted it.
Chinese builders (@laogua_daily, @xiaoyaoshowmore) described a cron task whose job was to report new posts. It saw the text “Run hermes update” and ran it. Their point: reading and acting collapsed into one step. Engagement is low; the failure mode is not.
@HermesAgentTips reported live voice transcription in Hermes Agent so speech is visible in the chat instead of a blank wait. Follow-up: available now via `hermes update` on the CLI only. Desktop app gets it in the next bundled release. Docker users stay on the stable tag; Termux and desktop wait for that bundle.
Nous Research announced Hermes Agent on the new ASUS ProArt RTX Spark PCs (P16 and P14). The stack pairs Hermes with a dedicated MuseTree skill (FLUX image, Wan video) and ComfyUI control, plus local models on up to 128 GB unified memory. ASUS press (Oct 7–8) says pre-orders are open, first US units Oct 16, NVIDIA RTX Spark up to ~1 PFLOP FP4, and quotes Dillon Rolnick on keeping workflows on-device. Teknium clarified related creative tooling is a catalog plugin, not something that ships inside Hermes itself.
OpenClaw DevRel McBain (@heyneighbor) posted a ~2.8 minute video covering recent work and the next stretch. Chapters he listed: new models (0:11), multiplayer (0:17), multiplayer in action (0:44), interactive dashboards (1:20), memory and skills (1:57), meetings and voice (2:08), easier Mac setup (2:20), what’s next (2:32). Official @openclaw quote-posted it.
David Rodecker reported the OpenClaw Foundation as the winning applicant for `.claw` in ICANN’s 2026 round, naming @davemorin, @steipete, and @openclaw. Steinberger confirmed: “WE GOT IT! .claw incoming!” and added that every agent should have a domain that is easy to reach. He answered a timing question with “Q4.”
@HermesAgentTips said Step 5 Preview was added free for a week in the Nous Portal model list. They used Hermes Agent with that model as the main brain to generate a website and posted a short screen recording.
@jessegenet described the cost of a poly-agent setup: remembering which agent and thread has the context. She is testing another product heavily but still runs a Nous Hermes agent team daily. Separate from that, @niyogi noted a VPS Hermes install that was told to install tooling and expose MCP itself.
A user who hit an update crash loop says maintainers picked up the report and expect fixes in 2026.9.33 by 12 Oct. Until then, upgrading from 7.35 may still crash-loop. Not an official changelog; treat as a hold-before-upgrade note.
Nous announced a $90M Series B led by Robot Ventures, with Nvidia, Microsoft M12, Samsung, Union Square Ventures, Y Combinator, Menlo Ventures, and others. Valuation is $1.5B. Capital is aimed at Hermes Agent enterprise controls (SSO, permissions, workspace isolation, audit, cost management, Slack/Teams) and a mobile app. Founder note says Hermes has been cloned 24M+ times and, by internal estimate, drives about 2.5% of global token usage. TechCrunch reports ~$36M annualized revenue by mid-September and a company expectation of passing $100M before year-end (WSJ figure, not independently checked here). WSJ, via Yuliya Chernova, also says Hermes overtook OpenClaw and other open-source assistants by usage on OpenRouter.
@arach describes hermes-scout, a plugin that gives Hermes a line into Scout so it can ask Claude Code, Codex, and similar coding agents to do work in repos and read the result back. Hermes itself stays usable without the plugin.
ASUS opened pre-order for ProArt P16/P14 RTX Spark Windows PCs (NVIDIA RTX Spark, up to 1 petaflop local AI, 128GB unified memory) bundled with ComfyUI and Nous Hermes Agent.
@fullctx shipped a native Tern frontend that keeps the same `hermes` command but streams tools and subagents live, writes replies in drying gold ink, and ends turns with a delivered mark. Follow-up PRs on 7 Oct add keyboard answers for approvals and clarifications, queue-on-busy so Enter does not eat a message, in-pane `/help /clear /cost /retry /doctor`, a 7-level thinking-effort ring, and honest truncation.
Operator note on the skill-learning loop: one bad instruction can ride into every later task that loads that skill. Gate it with `/skills approval on`, then `/skills pending`, `/skills diff <id>`, `/skills approve <id>`, `/skills reject <id>`. The gate covers skill writes including background learning.
Hermes Wingtips #93: the agent can search for a skill and run the install itself. Example prompt: “find a skill for flashcards and install it.” It picks a match, runs the security scan, then tells you what to say to use the skill. Aimed at fresh profiles and at people who do not know the skill name or the install command.
Luke The Dev’s digest of 142 PRs merged 5 Oct. Operator set: Desktop shows a progress chip before a subscription hits its usage wall, and rate-limited providers say why and until when in model pickers. Settings gets one Plugins pane. Plugin skills appear as `/plugin:skill` in TUI, Desktop, and CLI and can be stacked. `hermes chat -q` no longer hangs on approval, memory, MCP, or vault prompts. `hermes update` no longer stalls on a 20-minute git repack. Gateway reloads a platform whose plugin failed at startup without a full restart; Telegram flushes messages held during a disconnect. Dashboard login caps request bodies, bounds login values, and ignores spoofed `X-Forwarded-For`. Catalog memory providers (hindsight, mnemosyne) stop failing on every start, and installing one actually activates it. Memory writes on hand-formatted files fail with a fix hint instead of a silent miss. Opus 5.5 keeps thinking on; Anthropic summarized thinking is no longer promoted to the final answer. New hooks for human-input requests, post-gateway admission, and which cron run you are in.
Nous Research launched Hermes Index, a mean of four suites run inside the same Hermes Agent harness with reasoning set high where the model offers it. Suites: new Hermes Bench (150 tasks, 25 categories: skills, research, diagrams and creation, memory, tool use, safety), Terminal-Bench 4.0, Terminal-Bench Science, and SkillsBench. Graders look at workspace files, task state, and tool use. Portal: https://portal.nousresearch.com/bench. Official first board: Claude Opus 5.5 at 63.31 and $4.99/task, GPT 6 Astra at 56.25 and $11.61, Sonnet 5.5 at 53.14 and $2.82. Cheap end from the same thread: DeepSeek V4.1 Flash 36.91 at $0.26, Ling 3.0 Flash 21.56 at $0.054. Community recaps also put GPT 6 Sol near 44.10 / $2.23 and Grok 4.7 near 39.32 / $10.77, and flag some Opus and Sonnet cells as provisional. Treat those extras as secondary until checked on the portal.
freeCodeCamp posted a course on OpenClaw as an agent and messaging gateway (Discord and similar), plus the host-protection practices that matter once the agent can run workflows. High save rate relative to likes (about 443 likes, 403 bookmarks, 25k views at capture).
Recap of ThreatDown’s CARBONATO report (dated in the post as 22 Sep). Claim: bots hit unauthenticated Docker API on port 2375, drop a stock Hermes Agent, and replace the instruction file with a 39-line brief that harvests AI API keys before SSH keys or databases. Named targets include OpenAI, Anthropic, Google, Ollama, and vLLM. Stolen keys were pointed at an attacker-run API that advertised 12 models and served 27. Advice in the post: do not publish Docker’s API without auth; rotate AI keys like bank credentials. Primary write-up is the vendor report, not this post.
Update only. The docs page announced yesterday (326 stories) now has 45 more, mostly work use. Examples called out: Razorpay giving 220+ employees their own always-on agent; a pool company flagging bad technician reports while the route is live; a physician saving 30–45 minutes of documentation per case; a Swiss agency drafting invoices a human still sends; payment reconciliation for a local business; a bookstore running one agent for staff and one for customers. Page: https://hermes-agent.nousresearch.com/docs/user-stories. Nous reposted the thread early 7 Oct UTC.
Nous Research published a catalog of 326 user stories at hermes-agent.nousresearch.com/docs/user-stories. Tiles cover 15 categories (dev workflow 77, personal assistant 49, integrations 32, and others) and link back to the original X, Reddit, GitHub, Discord, YouTube, or blog post. Strong examples on the page: zero-token cron via `--script --no-agent`, a local email gatekeeper that hides 2FA and banking mail, and `/rollback` after a destructive git reset.
Teknium pointed at the plugins catalog (https://hermes-agent.nousresearch.com/docs/plugins): nearly 500 community mods for capabilities, model providers, memory, and UI. He later clarified only 8 are official/tested. Same window includes a pluggable computer-use backend so TryCUA can be swapped: https://github.com/NousResearch/hermes-agent/pull/133556
Paperclip v2026.1001.0: hardened native runner and chat recovery, agent personas across the app, agents as GitHub PR review bots, Railway connector, queued approvals. 77 commits, 8 contributors. Release: https://github.com/paperclipai/paperclip/releases/tag/v2026.1001.0. Related: NetworkChuck’s AI IT department video using Claude Code, Hermes, Codex, and local models under Paperclip roles and human approvals (https://www.youtube.com/watch?v=7RVf25Rg0Mc).
ClawStation (hosts OpenClaw) reported a real ops trap: npm treats a Node engine mismatch as a warning (EBADENGINE), installs the release anyway, then the binary refuses to start. Symptom is a gateway restart loop, empty journal, 502 at the proxy. Fix is raise Node to the release engines field, not downgrade OpenClaw.
Teknium published a catalog of open-platform hardware and devices Hermes (or any agent) can build on, integrate with, or live inside, with DIY project notes per item. URL: https://teknium.io/hermes-devices/
Priors shipped a skill that lets the assistant check another agent’s repayment record before paying or trusting it, and pay, borrow, or repay inside operator caps. Install: `hermes skills install priors-agents/agent001/skills/priors`. OpenClaw skill is in their hub. Rails are unsecured USDG lines and x402 on Robinhood Chain.
Dan Kornas shipped LaPis, a local persistent memory layer for Pi, Claude Code, Hermes Agent, and MCP clients. Stores decisions, bug fixes, constraints, patterns, and doc discoveries in SQLite. Session start loads relevant memories. Code and Markdown are indexed. Memories tied to changed code get a trust flag. MIT, no hosted memory service required. Link was placed in the reply, not the main post.
@downloadlos posted "I just broke up with my instinct" with two screenshots (about 224 likes, 29.6k views). Nous replied: this is why you need an agent that is truly yours and cannot switch models on you without warning. A reply in that thread said Hermes runs their life and they would not touch Instinct given the privacy controls. Parent post does not spell out the model change in text; the claim lives in the screenshots and the Nous quote.
TelepathicPug's team ("The Stigmergists") at the Oct 3–4 AI Village x Grove Research swarm hackathon built a custom Hermes plugin that crawls, collects, and grades public traces of AI swarm activity. Repo also has stylometry, morphology, taxonomy, and coordination notes. Video on the post.
Peter Steinberger posted that people are misinformed and pointed at https://www.openclaw.org: OpenClaw is an independent non-profit. In the same window he said the foundation does not need to share search results with a vendor or push a specific inference platform, and that there is no telemetry unless you opt in to help. He also confirmed one gateway can handle multiple agents, with isolation depending on security requirements.
Luke's Oct 4 digest covers 144 merges from October 3. Operator set: Desktop treats Ultrafast as its own speed tier; keep-awake can hold the machine only while a turn is running; bots keep each backend's identity across cold start when two backends share a profile name. Third-party plugins can run in a per-profile plugin host (`plugins.isolation: host`); portable plugins can require an NVIDIA GPU. Catalog delists 11 plugins that override Hermes core or Desktop, plus herdr-auto-reconcile and localsend (repos deleted). Dead MCP servers lose leftover processes even if they ignore SIGTERM. Profile export no longer ships `.ssh`, `.aws`, `.gnupg`, `.kube`, `.envrc`, or credential stores. Browser pages can no longer spend the subscription proxy credential; API keys stay on their exact origin. Telegram no longer stalls other chats on a slow one. Discord allowlist ignores nicknames and revoked IDs. Email trusts only the topmost auth header and requires an authenticated From under open access. Compaction no longer stubs a refused write as written, or summarizes a denied tool as if it ran.
Nous confirmed Hermes has iMessage and linked the Photon messaging guide. Separately, Nick Taylor (Pomerium) is speaking at MCP Dev Summit Toronto on 2026-10-05 on exposing OpenClaw as a remote MCP server with Pomerium as the MCP gateway. Talk is scheduled, not a shipped OpenClaw feature.
Teknium posted that the Hermes plugin catalog has been growing fast: nearly 500 mods covering capabilities, model providers, memory, and visual components. Link is the docs catalog index. Same thread, Bruno Oliveira pointed at Prompt Studio (already in catalog): F4 on a rough request, a few guided questions, a prompt tuned for Opus, Sonnet, or Astra.
Tony Simons published a one-page cheat sheet, docs checked 2026-10-03. SOUL.md is personality and tone. USER.md is the user’s profile and preferences. MEMORY.md is facts the agent has learned. AGENTS.md / .hermes.md is project instructions. SKILL.md is a reusable procedure. Session history is prior chats, not a notes file. Memory saves persist; the prompt snapshot refreshes in a new session. Passwords and API keys stay out of the notes.
Teknium posted a “Discord Setup, Checked For You” card: a wrong token is re-asked and never saved; Message Content Intent off gets a direct link to the toggle, then a re-check; the invite link prints with the right permissions; the owner is allowlisted without Developer Mode; a Discord section landed in the .env template. Same evening, Nous posted Linus Sebastian talking through a Hermes setup on the latest podcast (not sponsored, ~2 min clip). A reply in that thread said the Discord piece took three people a day. Nous answered that the Discord team has not taken their setup-stability suggestions, and that they had reached out so it should not take three days.
Official v2026.9.8 (43 PRs, 8 contributors named on the post; docs say the same scale). Delegated agent results return to the conversation that asked, once: inline if settled, or a single later delivery. Internal sessions no longer finish on NO_REPLY. Codex catalog processes hold fewer in-memory shell keys. Windows updates retry file locks; macOS npm updates accept aliased install paths. Docs are explicit that GPT-6.1 Sol was prepared (1.05M context, 128k output, 272k active input budget in OpenClaw) and was not complete in this release. Hannes Rudolph confirmed the miss: the bundled Codex CLI is 0.158.0 and needs a newer build. He posted three workarounds (updated ChatGPT app for computer use, force the OpenClaw runtime, or pin a 0.159.1+ Codex binary) and said 2026.9.9 is the fix. Not shipped by end of day.
Maurice (KI-Vater) posted a before/after of Hermes Plugin Guard v0.2.1 on a synthetic plugin. Removing `verify=False` clears the TLS warning. Outbound-call hints stay on purpose: a network call needs review and is not auto-malicious. The scan does not import or run the target. He is explicit that a pass is not a safety guarantee, and asked for real rule-id examples of useful vs noisy warnings.
Luke’s Oct 3 digest covers 175 merges from Oct 2. Hermes Release Watch pulled the operator set: catalog adds include Gmail, WhatsApp, Apple Health, Agent Builder, and several memory providers. Supermemory and Honcho install from the catalog; core no longer bundles Honcho. Memory-provider Python deps migrate on Desktop, the gateway, and scripted updates without a terminal. Plugin-guard denylist noise no longer blocks or prompts. Local MCP and plugin tools in one tool_call run instead of being rejected. SMS, email, and WeCom deliver long replies and cron output in full. Desktop skips rebuilds when HEAD did not move. Windows tray restore no longer leaves the window dead to input.
Peter Steinberger asked publicly for a Google contact after OpenClaw’s Android app sat in Play review for more than a week. Sundar Pichai replied “Ack, will follow up.” Sameer Samat, President of the Android Ecosystem, followed with “Sorry for that. Will get back to you quickly.” No ship confirmation by end of day. The app is still in review.
Tony Simons flagged the new catalog plugin `0bull`. Hermes can see the screen, read on-screen text, tap, type, run device commands, and execute macros on a real iPhone, not a simulator. He also says it can publish through the TikTok and YouTube Shorts apps on the device. Install path given: `hermes plugins install 0bull`.
BkashJosi posted a review prompt for Omarchy plugins: check the machine’s Omarchy version and dependencies, explain what the plugin adds, show install steps, include verification and removal, and wait for approval. Suggested starters: GitHub notifications, a calendar, Radio Atlas. Requires Hermes on the Omarchy machine so it can read the real setup.
Witcheer walked the Desktop path: Capabilities > Connectors, search the catalog (DeepWiki in the clip), Install, read the tool card, then ask for it by name in a new chat. Hermes Release Watch amplified it. DeepWiki is the example: questions against any public GitHub repo.
Aria Workspaces moved container Hermes to runtime 0.21.5 with plugin compatibility fixes, and OpenClaw to 2026.9.7 with gateway updates and config migration. Both upgrades are opt-in, take a state backup, and roll back on failure. Native Hermes stays on the official desktop-compatible release; 0.21.5 applies to container installs. Omarchy install flow moved to 4.0.4 with a whole-machine backup before native upgrades.
Luke’s Oct 2 digest lists 182 merged PRs. Hermes Release Watch pulled the three that change daily use: Codex OAuth can select `gpt-6.1-sol-900k` for a larger window; registered Nous accounts get managed Perplexity Fast Search without paid credits (guests stay on the keyless ring); kanban workers can reschedule their own cards with `kanban_schedule`. Security items in the same batch: the plugin catalog refuses plugins that override Hermes core or Desktop at runtime; X and Instagram embeds no longer run vendor scripts in the Desktop window; automatic git worktrees no longer run a cloned repo’s filters or hooks; at-rest secrets hardened in browser profiles, media caches, and the in-flight journal.
Patrick Erichsen (OpenClaw MTS) said 50+ plugins from Notion, Canva, Dropbox, and others landed. Install is from chat. Plugins hot-reload, so the gateway does not restart.
Nous Research confirmed multi-account Google handling: one install can authenticate several Google accounts, tokens stored per email, selection via `--account`, an environment variable, or a default-account setting. Multi-calendar access inside an account is supported (search, read, send, create events) after a Google Cloud project and OAuth credentials.
OpenClaw wired Tencent Hunyuan’s AI-Infra-Guard into ClawScan, the CLI that reviews every skill and plugin uploaded to ClawHub. ClawScan now runs AIG and NVIDIA SkillSpector independently, then an AI judge reads both reports plus the uploaded files. AIG covers nine risk classes, including instruction hijacking, memory poisoning, remote payload execution, unauthorized access, persistence, and insecure dependencies. On a 556-case SkillTrustBench subset, the combined system matched 86.9% of labels and correctly classified 98.6% of malicious cases. Disagreements and confirmed false positives go back to Tencent as regression cases.
@rlaope shipped a same-day thread on oh-my-hermes (https://github.com/rlaope/oh-my-hermes). It sits beside Hermes and does not patch it. Requests are scored into lanes (rename vs deep search), each lane has its own model chain with fallthrough, prompts are tuned per model and kept only while a benchmark says they help. Author claims the same coding tasks on GPT-6 Astra dropped from $4.29 / 23 min to $0.66 / 5 min. Also ships a cost HUD, verification gates, project memory, and 100+ workflow skills. Earlier post cites 3,000+ GitHub stars.
Teknium: you can now add full new languages across all Hermes surfaces via plugin, on top of the 16 already supported.
Paperclip highlighted a NetworkChuck build: an AI IT department with Claude Code, Hermes, Codex, and local models as assigned roles, with tasks and human approvals. Video: https://www.youtube.com/watch?v=7RVf25Rg0Mc
Teknium is back from vacation and says `hermes update` should now be about 4x faster for everyone. Nous Research replied in-thread. Community reaction was immediate (people running the update, HermesWatcher amplifying the desktop build speedup).
Official release on 2026-09-30. Snappier under load, smoother long chats, verified backups for more agent databases, automatic DB rollback if an eligible update fails before the new version starts, plus fixes for upgrades from 2026.9.5. Adds OpenAI Agents API and Sign in with ChatGPT (Beta), better Apple chat, and restart recovery. Scale: 2,818 PRs, 344 contributors. Docs: https://docs.openclaw.ai/releases/2026.9.7
Nous Research: “It wouldn’t be a very open agent if we barred people from using their Codex subs in it.” Direct product stance, not a rumor.
A community PR from @HermesAgentTips is merged: you can favourite a model in Hermes Desktop so the picker stops being a scroll hunt. @imbabybrooklyn posted the screenshot; Tips confirmed the merge.
@PaperclipShips: each agent now gets a personal directory that survives tasks and sessions (not just AGENTS.md). Completed Agent Chat handoffs queue back into the originating chat across restarts. Transport foundation landed so richer runners (Cursor, Copilot CLI, Pi — pending qualification) keep permissions, prompts, and in-flight state across reconnects. Community notes Paperclip around 95k GitHub stars after a ~13k-star day.
Official @openclaw posted Episode 12 of The ClawCast. @hrudolph and @Pat_Erichsen on decision models with TypeSafe AI; Kevin from OpenAI joins later on agents at work. Follows yesterday’s OpenClaw Enterprise launch.
@BkashJosi walked Desktop users through routing work across a multi-agent team: Bots tab to jump conversations; profile selector to stay inside one agent’s work. Research → researcher, code → developer, draft → content. Video of his own team.
@witcheer (Nous community) showed the Desktop `@` menu: type `@`, pick `@ url:`, paste a link, and Hermes fetches the page into the message before the agent reads it. Same menu attaches files, folders, git diffs, and recent commits. @HermesWatcher amplified it.
@HermesWatcher published a ranked list of the most-starred skills people actually drop into Hermes: Superpowers, Anthropic Skills, UI UX Pro Max, Graphify, Claude-Mem, Archify, Last30Days, OpenMontage, Humanizer, i-have-adhd, Marketing Skills, Planning With Files, SkillSpector, Skill Seekers, BrowserAct. ~62k views, heavy bookmarking.
@fullctx posted a short, high-res “QA with Hermes Desktop” clip (~10k views). Not a feature announcement — a visual of Desktop being used as a live QA harness.
Nous Research partnered with OpenAI so users can log into Nous Portal with ChatGPT and spend their existing Plus/Pro allowance inside Hermes Agent. Caps and visibility live in ChatGPT settings (Account Settings > Linked Accounts). No separate API key. Official video + ~149k views on the announcement.
OpenAI launched Dots (GPT-6 Astra): always-on personal agents with a cloud computer, 4,000+ app connections, and explicit permission boundaries. Community immediately framed it as the cloud, low-friction cousin of OpenClaw and Hermes — and as a possible replacement for self-hosted setups. Sam Altman: “Dots are here.” Official OpenAI video crossed 9M views.
The OpenClaw Foundation open-sourced OpenClaw Enterprise (OCE): a free, MIT-licensed, self-hosted control plane for persistent agents. Multi-tenant admin, fine-grained permissions, sandbox isolation, auditing, replaceable models/harnesses. Built to run on your infra via Docker Compose/Kubernetes. Aimed at internal pilots ahead of a 1.0 later in 2026. Official post ~159k views.
GitHub advisory GHSA-47wq-cj9q-wpmp: Paperclip could mint agent API tokens across tenants because `/api/agents/:id/keys` lacked an `assertCompanyAccess` check.
@Chadventures_ taught a Hermes agent to expand one OA lead into a store catalog, then re-price the catalog against Amazon using sales, coupons, and gift cards. Result: 219 leads in → 528 buyable leads out.
Hermes on macOS will soon pull SMS OTP codes directly from Messages for allowlisted sites. Password from vault → site sends text → Hermes reads the fresh code and continues. No human prompt, no OTP leaked into the model context.
Clawboo (open-source, MIT, self-hosted) launched with a live demo showing an OpenClaw agent acting as project manager: assigning a ticket to Hermes, Claude Code reviewing, Codex updating shared memory. One board, one chat, one memory across heterogeneous runtimes (OpenClaw, Hermes, Claude Code, Codex, native). `npx clawboo` gets you running; 400+ agents / 80+ teams available.
Tony Simons’ Hermes Fleet iOS app approved for external TestFlight. Manage bots, chats, groups, Kanban, artifacts, cron, and remote Hermes machines from phone.
ThaiCERT reports Carbonato botnet scans for open Docker API (port 2375), spins containers, and installs Hermes Agent to control hosts. No root required on host in basic case; privileged/volume mounts escalate risk.
Structured open-source book (CC BY-NC-SA) walking founders through building and governing companies where agents are the workforce, centered on Paperclip orchestration. Covers org design, kill switches, first 48 hours, scaling.
Patch release with ~460 merged PRs since v0.21.4. Improvements to Desktop plugin SDK, new interface mode, Connectors page, and expanded model support (GPT-6, Claude Opus 5.5). Update via `hermes update`.
One free NVIDIA API key unlocks several strong Chinese models as Hermes brains. Side-by-side comparison of reasoning depth on the same drawing task shows clear differences. Full setup article linked.
Enable `/memory approval on` so every proposed long-term memory save requires explicit approve/reject via `/memory pending`. Default is off in current docs (v0.21.5).
Community highlight of the official Home Dashboard plugin. Fully customizable widgets for live agent status, token usage, cron jobs, system stats, sessions, notes, and project tracking. Users are turning it into personal AI mission control by editing the plugin itself.
HermesWatcher clips Tommy (Nous) drawing the line between chatbot and agent: autonomy — booking reservations, building dashboards from data, running full CRM workflows, end-to-end creative pipelines. Nous already uses it this way internally.
Community PR (#93508) adds `hermes webapp` — serves the full Desktop experience (chat, files, Git, persistent terminal, remote sign-in) from any Hermes host. Teknium replied “Coming soon to all!” High engagement (1k+ likes on the original).
New native mode keeps the modern TUI inside the normal terminal instead of an alternate full-screen buffer. Native scrollback, output remains after exit, better terminal-first feel. Video walkthrough available.
Free open-source Chromium extension that turns Hermes into a side-panel copilot. Reads the active tab, injects page context into a live Hermes session (local gateway, Cloud, or self-hosted), and supports full tools/skills/memory. Can also interact with the page.
Upcoming Goal Mode turns delegated child agents into persistent workers. A separate judge evaluates progress; the same child continues with full context until the goal is met or budget is exhausted. Moves beyond one-shot delegated prompts.
Practical architecture for a 23-agent fleet with explicit roles, routing, human approval gates, and clear boundaries across Hermes, Grok Bots, and OpenClaw. Focus on lanes over model hype.
Script-only / no_agent cron jobs were running correctly but invisible in Desktop (blank description, no history). Update now shows the script, Script badge, fire times, and printed output by falling back to cron docs when SessionDB rows are absent.
OpenClaw shipped improved plugin loading with hot reloading. Demo video shows installing Apple PIM live, building plugins on the fly, and explains what developers need to change. Session can “cook” while plugins update without full restarts.
Microsoft announced Autopilot (formerly Scout), a persistent always-on personal agent for Microsoft 365, explicitly built on OpenClaw. The OpenClaw Foundation highlighted that Microsoft contributors (led by Omar Shahine) have pushed policy conformance, native Windows support, and reliability improvements back upstream. Blog details the bidirectional collaboration.
Hermes now supports `/branch` to fork a conversation into a parallel path (or dedicated thread in Discord/Telegram/Slack/Matrix) while preserving original context. Useful for exploring alternatives without polluting the main thread.
Detailed four-step walkthrough from HermesWatcher on turning a Hermes agent into a reliable autonomous worker using cron + Skills + progress files. Example: release-scout job that checks GitHub releases every six hours and produces sourced briefs without human intervention.
Nous Research announced that web search in Hermes Agent is now fast and free. Perplexity built Fast Search specifically for agents (p50 latency 160 ms, p95 230 ms). It is the default for all Nous Portal subscribers across tiers—no separate API key or extra bill required.
Detailed operational write-up of migrating from OpenClaw to Hermes, including aligning 47 cron jobs, verifying old processes/ports are gone, and the practical “scary” steps. Community commentary continues to position Hermes as a more polished evolution of the OpenClaw-style agent pattern.
Meta’s Muse agent is described as built from scratch but heavily inspired by OpenClaw. Concurrent community reports note that requesting a file export from Muse returned a 6.8 GB dump containing Linux root filesystem, SSH keys, and internal docs. Parallel discussion of OpenClaw’s own history with overly broad export behavior and a recent large deletion of test code (~400k lines) with minimal coverage impact.
Nous Research shipped Bot Screen for Hermes Desktop. You can now stream any bot’s screen (browser, terminal, apps) in real time, watch the agent work live, take control for 2FA/CAPTCHA/login steps, then hand the session back seamlessly. Works for remote/always-on gateways. Full docs and a detailed 5-min walkthrough by @tonbistudio available.
@ktrap used a Hermes agent to scan 122k+ on-chain CREDITS artworks, rank candidate letter shapes, generate review sheets, and assist human curation to complete a pixel font (added 7 letters to reach 21/26). Machine surfaced possibilities; human filtered for legibility.
@witcheer released a ready-to-paste prompt that makes Hermes audit its own shaping files (SOUL.md, USER.md, MEMORY.md, project AGENTS.md). The agent checks each line against the actual machine and project state, flags stale/duplicate/contradictory instructions, and proposes changes without applying them.
Official OpenClaw release 2026.9.6. Highlights include support for Opus 5.5 / GPT-6 Sol/Luna / Grok 4.7, managed updates, restart recovery, 30-day usage stats, GitHub reader, remote files/memory/skills, live meeting notes, and decision models (Jev + ONNX). 2,614 PRs from 351 contributors.
Hermes can expose any configured agent (skills, memory, tools, files) via an OpenAI-compatible endpoint. Apps such as Codex, Cursor, Cline, Open WebUI and AnythingLLM can now call into a running Hermes agent—including one on a remote VPS—without leaving their UI. Recent update properly streams model reasoning so the client app can show thinking in real time.
Nat Friedman (Meta) stated publicly that Muse was built from scratch but is “definitely heavily inspired as a product by openclaw.” After using OpenClaw in January the team bought hundreds of Mac minis and fell in love with the pattern. Goal was a safer, scalable version of the same idea.
DigitalOcean launched Managed Agents in public preview. Users can run Hermes (alongside Claude Code, Codex, OpenCode, LangGraph or custom containers) inside isolated microVMs that pause when idle. Tools are exposed through a single governed Action Gateway / MCP endpoint. One cloud, one bill, persistent state across sessions.
ZeroSignal is now in the official Hermes Plugin Catalog. It provides encrypted, privacy-preserving inference across 50+ models served by independent operators. No API keys, no accounts, no centralized prompt history. Context is encrypted on-device before leaving the machine.
Malicious `.git/config` with `core.fsmonitor` can fire when the agent refreshes git status after a user message. Full environment (including provider API keys) is exposed. Affects builds prior to commit f6234d0. Do not open untrusted repo archives. The 0.21.0 tag alone is not proof of the fix.
Paperclip now lets agents search and read team knowledge in Notion. Separately, the Connectors system lets operators set per-tool permissions: Off / Ask first / Allowed. Agents get only the access required for real work.
Nothing matches. Try a builder’s handle or a harness name.
