Hermes Agent 0.22.0: Threshold release with local llama.cpp and guided setup
On October 14, 2026, Hermes Agent 0.22.0 was released — "The Threshold Release." It is an agent environment from Nous Research with a built-in learning loop: the agent creates skills from experience, improves them through use, and accumulates memory about the user between sessions. Since the last review, patches 0.21.3—0.21.7 came out, and 0.22.0 consolidates them into one major stable release.
What's new
- First launch with guided setup. On a clean machine, the desktop app opens a short setup chat on a free trial model — a newcomer starts talking to the agent before they even have an API key.
- Local models in one click. The built-in llama.cpp runtime downloads the model, manages memory, and spins up the server on its own — running your own model on your own hardware without manual setup.
- Stable releases by default. Source installs now update to tagged releases rather than every commit on main. The channel can be switched in the desktop settings; for those who want to live on the bleeding edge, hermes update --set-channel main remains.
- Signed packages and Microsoft Store. Desktop builds are signed, including the store version, and a single package manager underlies everything.
- Reasoning survives resume. The model's signed "thoughts" are reproduced after hermes chat --resume and on the next gateway turn.
- A catalog of 600 plugins and a unified backend with a gateway for each profile.
What the platform does
Hermes Agent is an environment extended with skills and tools. It can spawn isolated subagents for parallel tasks, write Python scripts that call tools via RPC, connect any MCP servers, maintain agent-curated memory with periodic reminders and autonomous skill creation, and request human confirmation for commands. Voice and speech are claimed: transcription of voice messages and a unified gateway for Telegram, Discord, Slack, WhatsApp, Signal, and CLI.
Your own model instead of the cloud
Your own model connects via Ollama, llama.cpp, and vLLM. The documentation separately describes the nuances: llama.cpp needs the --jinja flag, vLLM needs --enable-auto-tool-choice --tool-call-parser hermes, otherwise tool calls arrive as text. The platform can manage the local llama.cpp server itself — from model download to memory. MIT license.
How it differs from its neighbors
Of the three neighbors in the registry, only Hermes Agent claims voice and speech in its documentation. Sandboxed code execution, which most neighbors have, is not mentioned in Hermes Agent's documentation. Custom tools and functions are claimed by all four platforms.
| Hermes Agent | DeerFlow | FrontierAgent | OpenBot | |
|---|---|---|---|---|
| Type | environment | environment | environment | environment |
| Language | Python | Python | Python | TypeScript |
| License | MIT | MIT | Apache-2.0 | MIT |
| GitHub stars | 254,468 | 82,638 | 3,903 | 5,112 |
| Releases in 90 days | 20 | 1 | — | 12 |
| Multiple agents in a team | yes | yes | yes | — |
| Custom tools and functions | yes | yes | yes | yes |
| MCP support | yes | yes | — | yes |
| Agent memory | yes | yes | — | yes |
| Sandboxed code execution | — | yes | yes | — |
| Browser automation | — | — | — | yes |
| Human confirmation | yes | — | yes | — |
| Graphs and state persistence | — | — | yes | — |
| Tracing and debugging | — | yes | — | — |
| Agent quality evaluation | — | — | yes | — |
| Voice and speech | yes | — | — | — |
| Own server and Docker | yes | yes | — | yes |
| Own model: Ollama | yes | — | — | — |
| Own model: llama.cpp | yes | — | — | — |
| Own model: vLLM | yes | yes | — | — |
| Own model: SGLang | — | — | yes | — |
| Own model: OpenAI-compatible API | — | yes | yes | yes |
Cross-checked against the projects' READMEs and documentation as of 10/17/2026: "yes" means the capability is claimed there, a dash means it is not mentioned (which does not mean the capability is absent). Stars and releases are per GitHub data.
It is worth a look for those who want to run an agent on their own hardware with minimal setup while chatting with it from messaging apps. Update to 0.22.0 if you are on source installs and are tired of updating to every commit. Following releases like this is convenient via the CPU3D news feed.
