Hermes Agent v0.22.0 adds one-click local models and 600 plugins, but turns on 90-day session deletion by default and needs a follow-up patch
Nous Research's "Threshold Release" is the project's biggest window yet (about 6,080 merged PRs by its own count). Install v0.22.2, not 0.22.0 or 0.22.1, and check sessions.auto_prune first.
Every release passes source CI, Docker, Nix, signed desktop builds and install/upgrade runs on every OS before it is published, and source installs now move from release to release by default instead of tracking every commit on main.
Comprehensive take
v0.22.0 is a genuine product step for Hermes: first-run setup, connectors, local models and a stable release channel remove the main reasons a newcomer would bounce. It also changes defaults that can bite existing users (automatic deletion of sessions inactive for 90 days, Python 3.14, MSIX-only Windows, Apple-Silicon-only macOS) and the 0.22.1 wrong-commit build shows the new release pipeline is not yet proven. Update to v0.22.2 deliberately, with auto-prune set the way you want it.
Use cases
Run an agent fully offline on your own GPU
The Local Models page installs llama.cpp for your hardware and only recommends builds that fit fully in GPU memory, with no account or API key.
Let the agent work in Linear or Notion
Connectors expose hosted apps through the Nous tool gateway, with multiple named accounts such as work and home.
Operate one gateway for several profiles
A single multiplexed gateway and desktop backend serves every profile, which cuts processes and token clashes for people who run several bots.
Specs
| Release date | v0.22.0: 14 Oct 2026 UTC; v0.22.2: 15 Oct 2026 source |
|---|---|
| Window since v0.21.0 (project-reported) | ~24,200 commits, ~6,080 merged PRs, ~7,180 issues closed, ~2,520 contributors source |
| Plugin catalog | 600 reviewed plugins, each pinned to a reviewed commit source |
| Session auto-prune default | On: ended sessions inactive for 90 days are deleted at startup; pinned and open sessions untouched source |
| Platform requirements | Python 3.14 (managed); Windows 11 22H2+ MSIX only; macOS Apple Silicon only in default builds source |
| Local runtime | llama.cpp b11370; CPU, CUDA, Metal, Vulkan, ROCm source |
| 0.22.1 incident | Built from the wrong commit; 0.22.2 known issue: 0.22.1-gateway sessions cannot be deleted source |
Insights
- Axion reading: Nous is moving Hermes from a developer tool toward a packaged desktop product, and the key-free first launch is the clearest signal.
- Axion caution: the commit and PR totals are the project's approximations and say more about release size than quality; the wrong-commit 0.22.1 is the better quality indicator.
- Axion caution: the 90-day auto-prune default is the change most likely to surprise existing users.
What shipped
Hermes Agent v0.22.0, which Nous Research calls "The Threshold Release", was published on 14 October 2026 (the notes are dated 13 October). By the project's own count it rolls up roughly 24,200 commits, 6,080 merged pull requests, 7,180 closed issues and 2,520 contributors since v0.21.0. Those are self-reported approximations, but they explain why the notes run to more than 118,000 characters. Two patch releases have followed: v0.22.1 was built from the wrong commit and shipped unreleased gateway work (the Microsoft Store app could stick on "Connecting"), and v0.22.2 on 15 October is the release 0.22.1 was meant to be, with a known issue that sessions created under the 0.22.1 gateway cannot be deleted. If you are choosing a version today, choose v0.22.2, not v0.22.0 or v0.22.1.
The five changes that matter
Setup without a key. A fresh desktop install opens a guided setup chat that runs on a free trial model, nous/welcome, once per OS user. It is desktop-only: the CLI, TUI and gateway get no free model, and existing installs do not see it.
Connectors. The agent can act in hosted apps such as Linear and Notion through the Nous tool gateway, with several named accounts per app and a new Connectors page that merges hosted apps with MCP servers. Hermes creates a free connectors-only identity on first use, so no Nous account is required to start. Anyone with data-handling rules should note that this routes through Nous infrastructure; the notes provide `tools.connectors.enabled: false` and `nous.guest: false` to turn it off.
One-click local models. Hermes installs and maintains llama.cpp (build b11370) for CPU, CUDA, Metal, Vulkan or ROCm, rates catalog models against your GPU and RAM, and recommends only models that fit fully in GPU memory. The catalog lists Qwen3.8 27B, Qwen3.8 Flash Next, Qwen3.6 35B-A3B and DeepSeek V4 Flash, plus any Hugging Face GGUF.
Stable channels and signed packages. Releases now pass source CI, Docker, Nix, signed desktop builds and install/upgrade runs on every OS before publication, and source installs follow releases by default rather than `main`. A new `hermes pm` package manager builds a new environment and swaps only after a check passes. The patch-release history above shows the pipeline is still settling.
Models and plugins. Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5, GPT-6 Astra, Sol and Luna, GPT-6.1 Sol and Gemini 4 Argon are selectable where providers offer them, and the plugin catalog has 600 reviewed entries, each pinned to a reviewed commit.
Read the upgrade notes before you update
Four defaults change behaviour, and one deletes data. `sessions.auto_prune` now defaults to on: ended sessions inactive for 90 days are deleted at startup, though pinned and open sessions are never touched; set it to false to keep everything. Chat idle and daily resets are retired, so conversations persist until `/new` or `/reset`. Managed installs run Python 3.14. Windows desktop is MSIX-only and requires Windows 11 22H2 or newer, and default macOS builds are Apple Silicon only, so Intel Macs and Windows 10 users are left out. Memory providers such as Honcho, Mem0 and Supermemory moved to catalog plugins, with data migrated automatically, and plugin authors lose the old compatibility import layer.
What we could not verify
We read the release notes, not the code, and ran none of this. The prompt-size claim (about 2.1k fewer tokens per request) and the local-model recommendations are the project's own statements with no published benchmark attached. Telemetry sharing is opt-in and both switches default to off, per the notes.
Should you switch
If you run Hermes on a laptop with a GPU, v0.22.2 is worth installing for local models and for the gateway-after-reboot fix. If you run a long-lived gateway or any session archive you care about, set auto-prune explicitly before updating. If you are on Windows 10 or an Intel Mac, stay on your current build.