Nous Research has released Hermes Agent v0.22.0, "The Threshold Release", which its own release notes describe as the largest window the project has shipped. The numbers, as stated by the project: about 24,200 commits, 6,080 merged pull requests, 7,180 closed issues and roughly 2,520 contributors since v0.21.0, about 2.5 times the previous cycle. Release date: October 13, 2026.
The theme is getting a new user through the door. A fresh install now sets itself up on a free model, apps connect without an account, open models run locally in one click, and installs follow tested releases instead of the moving main branch. Below is what changed, what to check before upgrading, and how it fits with the Hermes coverage we have published since the bot mode release in August.
At a glance
| Feature | What it does | Who benefits |
|---|---|---|
| Guided first launch | Setup chat on the free nous/welcome trial model; no key needed | New desktop users |
| Connectors | Act inside hosted apps such as Linear and Notion via the Nous tool gateway | Anyone using SaaS tools |
| Local models | One-click llama.cpp runtime, hardware-aware model picks | Privacy-minded and offline users |
| Stable channels | Tested SemVer releases by default; main is opt-in | Teams that need reliability |
| Signed desktop packages | Windows MSIX (incl. Microsoft Store), notarized macOS build | Non-developers |
| Plugin catalog | 600 reviewed plugins pinned to commits | Power users, plugin authors |
| Bot Screen | Watch a bot's desktop, take over for login or CAPTCHA | Gateway operators |
| Leaner code | Source cut from about 1.06M to 698K lines | Contributors |
A first launch that needs no API key
Open the desktop app on a machine that has never run Hermes and it starts a short setup chat rather than a settings screen. It asks for your name, an accent color and layout, which apps to connect and which plugins to install, then offers up to four first tasks. It runs on the free trial model nous/welcome, usable once per OS user, so no key is needed. The agent also gained a set_theme tool, so "make it violet, dark mode" works in a desktop chat.
The free tier has limits worth knowing: a first desktop install with no provider or Nous account gets the model; the CLI, TUI and gateway do not. The notes say free rate limits end the turn immediately with the reset time, and free-tier users see a Terms and Privacy notice. If you later sign in with hermes auth upgrade or /login, the free identity upgrades to a full Nous account and keeps your connected apps.
Connectors: the agent works inside your apps
Connectors let the agent act in hosted apps like Linear and Notion through the Nous tool gateway. You ask for an app, sign in once in the browser, and the tools work in the same turn. A free connectors-only identity is created the first time you use an app. The desktop gets a Connectors page that lists hosted apps and your MCP servers together, replacing the old MCP tab. You can connect several named accounts of one app, such as work and home, and manage them with /connectors or hermes connectors.
If you are new to the underlying protocol, our MCP explainer covers what an MCP server is and why one list for both kinds of tools is convenient. Note the trade-off: hosted connectors route through Nous's gateway, whereas a local MCP server stays on your machine. Turn connectors off with tools.connectors.enabled: false if that matters for your data.
Local models in one click
Hermes can now run open models entirely on your machine, with no account and no key. It installs and maintains llama.cpp with the right engine for your hardware: CPU, CUDA, Metal, Vulkan or ROCm. It rates each catalog model against your GPU and RAM and picks the highest-quality build that fits. Downloads pause and resume, the server starts and stops with Hermes, and engine updates are one click. The release notes say every backend runs llama.cpp build b11370.
The catalog includes Qwen3.8 27B, Qwen3.8 Flash Next with its multi-token-prediction speculative head, Qwen3.6 35B-A3B and DeepSeek V4 Flash, plus any GGUF from Hugging Face. Recommendations only include models that fit fully in GPU memory, and Qwen3.8 Flash Next is noted as fitting 128 GB machines. Find it under Settings, Providers, Local Models, or hermes model and then Local models in a terminal. For background see our posts on running a local agent on a Mac Studio and NVIDIA's PAIR with Hermes Desktop.
Stable releases and signed packages
Previously, source installs tracked every commit on main. Now Hermes ships tested SemVer releases. Each release passes source CI, Docker, Nix, signed desktop builds and install and upgrade runs on every OS before publication, and source installs follow release to release by default. To keep the old behavior, run hermes update --set-channel main, or use the switch in Desktop Settings under About. Moving to an older release is labelled a rollback and never happens without asking.
The desktop app is also now a sealed, signed package that carries its own Python and Node. Windows gets a signed MSIX for x64 and ARM64 through App Installer or the Microsoft Store, and macOS gets a signed, notarized Apple Silicon build. Underneath is a new package manager, hermes pm, that owns pinned, hash-checked tools and Python environments. Updates build a new environment and switch only after it passes a check, so a killed update cannot leave an install half-applied. On one host, garbage collection reclaimed environments from 92 GB to 18 GB, according to the notes.
Plugins, models and the agent loop
- Plugin catalog: 600 reviewed plugins, each pinned to a reviewed commit, installable from Desktop or the CLI. Rules keep listed plugins from patching Hermes internals or shipping unscannable binaries. Memory providers such as Honcho, Supermemory, Hindsight and Mem0 moved to catalog plugins, with automatic migration.
- Models: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, GPT-6 Astra, Sol and Luna, GPT-6.1 Sol and Gemini 4 Argon are selectable on Nous Portal, OpenRouter and native routes where offered. See our Astra vs Sol comparison if you are choosing.
- Reasoning replay: prior reasoning is replayed by default on providers that can read it, across resume, branches and compaction, and transient 503, 529 and timeouts are waited out in bounded cycles.
- Password-manager sign-in: the browser agent can log in using 1Password, Bitwarden or Hermes's encrypted vault without the model seeing the password; checkout card and address fills need confirmation.
- Bot Screen: a bot on a Linux gateway can have its own desktop streamed live; when it hits a login, 2FA or CAPTCHA you take over, then hand back. With Docker, SSH or Singularity backends, the bot's screen and browser run inside the sandbox.
- Subagent control: running subagents stay visible above the composer with live transcripts plus Steer and Stop, which builds on the manual subagent control shipped in September.
- Model picker: shows how much of a subscription remains and when a rate limit lifts, and lets you set reasoning effort with
/model X --reasoning high.
Upgrade notes that matter
The project lists several breaking changes. Read them before you run hermes update:
- Managed installs use Python 3.14; KittenTTS is unavailable there.
- Windows desktop is MSIX-only, requires Windows 11 22H2 or newer, and the NSIS and MSI installers are gone. Default macOS builds are Apple Silicon only.
- One gateway per machine is the default;
hermes updatefolds per-profile gateways into it when nothing conflicts. - Chats no longer reset on an idle or daily timer; use
/newor/reset. - Old sessions are pruned automatically:
sessions.auto_prunedefaults to on and deletes ended sessions inactive for 90 days. Pinned and open sessions are never touched. Set it to false to keep everything. - Plugin authors: the old compatibility import layer,
hermes syncand the bundled tirith scanner were removed; the task-list tool pluscronjobandprocesswere renamed, with old names still working.
The auto-prune default is the one most likely to surprise people. If you rely on old transcripts for audits, disable it before updating.
Security context
The intermediate v0.21.6 notes list fixes reported by Tenable Research for the dashboard: rate-limit bypass through a spoofed X-Forwarded-For header, unbounded audit-log writes, missing body-size limits on public auth routes, and a sign-in redirect that could allow session takeover. Other fixes covered untrusted git filter programs and an email-gateway sender spoof. Those are in the window rolled up by v0.22.0. If you run an agent gateway exposed to the internet, treat updates as security-relevant.
How it compares
Hermes is positioned against other open agent shells; our comparison of OpenAI Dots, Grok Bot, Meta Muse, OpenClaw and Hermes covers the field as of earlier this year. The v0.22.0 pitch is lower friction (free start, signed installers) and local-first options, plus a bigger plugin ecosystem than most rivals list. As with any release-notes-driven coverage, these are the project's claims; we have not benchmarked the new features.
Bottom line
If you tried Hermes and bounced on setup, this release targets exactly that. If you already run it, stage the update, read the upgrade notes, and decide which channel you want. The contributor and commit counts are the project's own, but the engineering direction, tested releases, signed packages and local models, is concrete and checkable in the changelog.
Details reflect the v0.22.0 release notes as of October 14, 2026.