Hermes Agent v0.22 adds local models, stable channels, plugin catalog

Hermes Agent is Nous Research's AI agent, which runs as a desktop app, in the terminal and as a gateway for chat bots. Version 0.22.0, which the team calls the Threshold Release, is mostly about making it easier to start. Its release notes say a fresh desktop install now opens on a short setup chat instead of a settings screen, running on a free trial model, so a newcomer can talk to the agent before owning an API key. The chat asks for a name, an accent color and a layout, which apps to connect and which plugins to install, then offers up to four suggested first tasks.

Three changes carry the rest of that theme. Connectors let the agent act inside hosted apps such as Linear and Notion through Nous's tool gateway: you sign in once in the browser and the tools work in the same turn, and Hermes creates a free connectors-only identity, so no Nous account is needed. Local models go the other way, keeping everything on your own computer with no account or key. Hermes installs and maintains llama.cpp, an open-source program for running models on ordinary hardware, picks the engine that suits your graphics card or processor, rates each catalog model against your GPU and memory, and chooses the best build that fits. The catalog includes Qwen and DeepSeek models plus any GGUF file, the common format for such local models, from Hugging Face.

The third change is stable release channels. Previously a source install tracked every commit on the main branch. Now, the team says, each release passes CI, Docker, Nix, signed desktop builds and install and upgrade runs on every operating system before publication, and installs move from release to release by default; anyone who wants the newest code can switch back to main. The desktop app also became a signed package that carries its own Python and Node, with a signed Windows build available through the Microsoft Store. A new package manager builds each update in a separate environment and switches only once it passes a check, so an interrupted update cannot leave an install half-applied. Around this the notes list a plugin catalog of 600 entries, each pinned to a reviewed commit, support for newly released models from Anthropic, OpenAI and Google, and browser logins filled from a password manager without the model seeing the password. GitHub cuts the v0.22.0 notes off partway through, so later sections were not read.

Two patches followed within two days. The v0.22.2 notes say v0.22.1 was built from the wrong commit, so it shipped unreleased gateway work and could leave the Microsoft Store app stuck on Connecting; 0.22.2 is the release 0.22.1 was meant to be, with one known issue: sessions that ran under the 0.22.1 gateway cannot be deleted. It also fixes a gateway that refused to start after a reboot, and long local-model turns that failed on laptops. The pull request for that fix says a press reviewer saw single queries to a 27-billion-parameter Qwen model take hours, because the agent's timers cut off a server still reading a long prompt or one paused by the laptop sleeping. The previous major version, v0.21.0, was tagged in early September.