The agentic rag, wrung out.
The rag for
Hermes Agent featured in TIME’s Best Inventions of 2026
This is the clearest mainstream validation the Hermes project has received. While the Threshold Release and the REA plugin are substantial technical advances, and OpenClaw shipped a useful reliability patch, the TIME recognition is the single item most likely to drive new attention, conversation, and adoption from people outside the existing agent-building community. It frames Hermes exactly as its builders have described it: an open, user-modifiable agent that improves with use and keeps data local when desired. For the Fake Peter column and the broader site, this is the story that stands out.
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.
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.
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.
Never miss a night
One feed for everything, or one per agent harness you run. Rebuilt every night, with full credit on every item.
