Hermes Desktop is the official desktop app for Hermes Agent, the MIT-licensed AI agent from Nous Research. It is the same agent as the hermes command line with a graphical workspace on top, free to download for Windows and Apple silicon Macs, and built from the CLI on Linux. It can run entirely on a local model through a built-in llama.cpp runtime, Ollama or LM Studio.
People looking for the app tend to ask the same things: what it is, which installer to pick, whether it replaces the terminal agent, how to point it at a local model, and what its plugins can do. This guide answers those from Nous Research’s documentation, the Hermes Agent repository and its release notes, and Ollama’s integration page, all read on 14 November 2026, when the newest stable release was Hermes Agent v0.23.1 (10 November).
One naming note first. Three projects use the name. This guide covers the official app, and the first section shows how to tell the other two apart.
What is Hermes Desktop?
Hermes Desktop is a native application for Hermes Agent, the agent that Nous Research publishes under the MIT licence. Its documentation is direct about what it is not: a separate product or a cut-down clone. The app uses the same agent core, configuration, API keys, sessions, skills and memory as the hermes command, so anything set up in a terminal is already there when the app opens.
It arrived with Hermes Agent v0.16.0, which the release notes date 5 June 2026 and name the Surface Release. The app’s code sits in the apps/desktop folder of the main repository, which had about 263,000 GitHub stars when we checked.
Technically it has three parts:
- An Electron shell that owns the window, file system and Git access.
- A React interface for chat, panes and settings.
- The agent itself, started as a headless
hermes serveprocess that the interface talks to over JSON-RPC and WebSocket.
On top of chat with streaming tool output, the app adds a file browser, side-by-side previews of web pages and files, voice dictation and read-aloud, a terminal pane, Git review with worktrees, scheduled jobs, profiles, and projects that group folders and sessions.
The other apps called Hermes Desktop
| Project | Maker | What it is |
|---|---|---|
| Hermes Desktop (this guide) | Nous Research | The official app, in NousResearch/hermes-agent under apps/desktop, MIT licence |
fathah/hermes-desktop, now branded Hermes One | Community | An MIT-licensed desktop companion that installs and configures Hermes Agent through the official install script; the repository dates from April 2026, before the official app |
dodo-reach/hermes-desktop | Community | A Mac tool that manages a Hermes install over SSH |
If a tutorial’s screenshots do not match your app, it was probably written for one of the community projects. The official downloads are on hermes-agent.nousresearch.com.
Hermes Desktop vs Hermes Agent
Hermes Agent is the agent. Hermes Desktop is one way to use it. The documentation lists several front ends that all talk to the same agent and share state:
| Front end | How you open it | Best for |
|---|---|---|
| Desktop app | The installed app, or hermes desktop | A graphical workspace, previews, one-click local models |
| CLI and TUI | hermes, hermes --tui | Servers, SSH sessions, scripts |
| Web dashboard | hermes dashboard | Browser-based administration |
| Webapp | hermes webapp | The desktop workspace in a browser tab, served by your own gateway; added in v0.23.0 and off by default |
| Messaging gateway | Configured under Messaging | Reaching the agent from Telegram, Discord, Slack and other chat apps |
Since v0.23.0, released on 5 November 2026, one gateway process on each machine owns every conversation and each front end attaches to it as a client. A chat opened in the app is the same chat in the terminal, with one history.
So the choice is about interface, not capability. Use the app when you want to see files, diffs and previews next to the conversation. Use the CLI on a headless server, then connect the app to it from Settings > Gateways; in that remote mode, the documentation notes, tools, terminal commands and file operations run on the remote host.
If you are still deciding between agents rather than interfaces, our comparison of Hermes Agent with Claude Code and OpenCode covers the coding side, and the Hermes and OpenClaw comparison covers personal assistants.
Hermes Desktop download for Windows, Mac and Linux
Nous Research’s download page listed these options on 14 November 2026:
| System | What the site offers | Notes from the documentation |
|---|---|---|
| macOS | Hermes-Setup.dmg, labelled macOS 12+ | Apple silicon only. No release builds an Intel bundle, and the installer reports that it is not supported on an Intel Mac |
| Windows | Microsoft Store package, or Hermes-Setup.exe, labelled Windows 10/11 | The signed MSIX package requires Windows 11 22H2 or later. Hermes-Setup.exe is a bootstrap installer that sets up a source installation instead |
| Linux | Install via terminal | Linux desktop packaging is disabled in the release workflow, so there is no AppImage, deb or rpm to download; the CLI builds the app locally |
The documentation separates two kinds of installer. A bundled package contains the agent, Python and prebuilt interfaces, and does not build anything on first launch. A Hermes-Setup bootstrap installer downloads a source installation and builds the desktop app on your machine, which takes longer and needs network access.
Install on Windows or an Apple silicon Mac
- Download the installer from the Hermes Desktop page on Nous Research’s site, not from a mirror.
- On macOS, open the DMG and copy the app to Applications. On Windows, install from the Microsoft Store or run the installer you downloaded.
- Start the app. A fresh install opens a short setup chat. Choose Skip setup or Choose provider later if you want to configure a local model before sending anything.
Install on Linux, Windows 10 or an Intel Mac
Install the command-line agent, then let it build the app:
# Linux, macOS and WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes desktop
# Windows (PowerShell)
iex (irm https://hermes-agent.nousresearch.com/install.ps1)
hermes desktop
hermes desktop installs the Node dependencies, builds the Electron app for your system and launches it against your existing configuration. On Linux it also writes a launcher entry so the app appears in the application menu. Read any install script before piping it into a shell, as you would for any tool that will run commands on your machine.
Install through Ollama
Ollama documents its own route, which installs Hermes if needed, asks you to pick a model and writes the configuration:
ollama launch hermes-desktop
ollama launch hermes-desktop --model <model>
Ollama’s page adds that on macOS and Windows the upstream installer is the recommended path, and that both routes share the same Hermes configuration, sessions, skills and memory.
Updating and uninstalling
The app checks for updates in the background and offers a one-click update. Store packages update through the Microsoft Store. To remove an app-managed install, open Settings > About > Danger zone, which offers three levels: the app only, the app and agent with your data kept, or everything.
Hermes Desktop with a local model
There are four routes to a local model. All of them share one rule: Hermes needs at least 64,000 tokens of context for agent use with tools, and its documentation says smaller windows are rejected at startup.
| Route | Where you set it | Default address | Context setting |
|---|---|---|---|
| Built-in Local Models | Settings > Providers > Local Models | Managed by the app | Automatic; every recommended model gets at least 64K |
| Ollama | Custom endpoint, or ollama launch hermes-desktop | http://localhost:11434/v1 | OLLAMA_CONTEXT_LENGTH=64000 or a Modelfile |
| LM Studio | The LM Studio provider | http://localhost:1234/v1 | Context Length in LM Studio, or lms load --context-length 64000 |
| Your own llama-server | Detected, or a custom endpoint | Port 8080 is the only one probed by default | -c 64000 or higher |
The built-in Local Models runtime
Since v0.22.0, released on 13 October 2026, the app can run models without any other software. The documentation gives four steps:
- Open Settings > Providers > Local Models, or choose Run models locally during onboarding.
- Click Install runtime. Hermes downloads a llama.cpp build for your hardware, a few hundred megabytes, and verifies it against pinned SHA-256 hashes.
- Pick a model from the catalogue and click Download.
- Click Use. New chats run on the local model.
Each catalogue row shows whether the model fits your GPU, spills into system RAM or is too large. Hermes picks the quantisation for you and never offers builds below 4-bit. It recommends the highest-quality model that it predicts will run fully in GPU or unified memory at 20 tokens per second or more. The docs give a rough guide: a GPU with 8 GB or more runs the small catalogue models, and 16 GB or more runs the 27B to 35B models at high quality. Idle models unload after 15 minutes. A search box below the catalogue covers the rest of Hugging Face, and Add model file links a GGUF you already have.
Supported back ends are CUDA, Vulkan or CPU on Windows and Linux, and Metal or CPU on macOS; on Linux, CUDA needs a CUDA 13 driver, and HIP/ROCm is an explicit choice on supported x64 machines. Our VRAM calculator estimates what a given card can hold, and the tool-calling model guide explains why some models suit agents better than others.
Ollama
Ollama’s default context is the usual stumbling block. Its documentation lists 4K tokens under 24 GiB of VRAM, 32K from 24 to 48 GiB, and 256K above that, and the first two are below what Hermes accepts. Raise it on the server:
OLLAMA_CONTEXT_LENGTH=64000 ollama serve
ollama ps # the CONTEXT column should show 64000
Then, in the app, add a custom endpoint under Settings > Providers > Custom Endpoints with the URL http://localhost:11434/v1, no API key, and the model name exactly as ollama list prints it. The context cannot be set through the OpenAI-compatible API, so the server-side setting is the one that counts. Tags ending in :cloud run on Ollama’s servers, not on your machine.
LM Studio
Start LM Studio’s server from its Developer tab or with lms server start, then pick the LM Studio provider in Hermes and accept http://localhost:1234/v1. Set the model’s context length to 64,000 or more in LM Studio and reload it. Hermes reads the context that LM Studio reports after loading.
Your own llama-server
If a llama-server is already running, Hermes detects it and uses it instead of starting its own. There is a catch for current llama.cpp builds: Hermes probes port 8080 by default, while llama-server has listened on port 9931 by default since build b11521 of 9 October 2026. Either start the server with --port 8080, or tell Hermes where it is in config.yaml:
providers:
llamacpp:
base_url: http://127.0.0.1:9931/v1
model: my-model
Keep parallel slots in mind. With -c 64000 -np 4, each slot gets 16K tokens, which is below the minimum. Our llama-server guide covers the flags, and the local agent walkthrough for OpenClaw and Hermes goes deeper on the command-line side.
Hermes Desktop plugins and skills
Three extension types are easy to confuse.
| Type | What it is | Where it lives | Runs as |
|---|---|---|---|
| Desktop plugin | A single ESM file that adds panes, pages, sidebar items, commands, keybindings or themes | $HERMES_HOME/desktop-plugins/<id>/plugin.js | Code inside the app, not sandboxed |
| Agent plugin | Python code that adds tools, hooks and integrations | ~/.hermes/plugins/<name>/ | Code inside the agent process |
| Skill | Instructions and files the agent loads for a task | Managed in the Skills pane | Text the agent follows |
Desktop plugins are managed under Capabilities > Plugins, and the app reloads a plugin file when it changes. The plugin catalogue is the curated source: the v0.23.0 release notes say it grew from about 600 to more than 1,050 entries, every entry is admitted through a maintainer-reviewed pull request, and each one pins an exact commit, so an author pushing new code does not change what the catalogue installs.
The documentation is unusually frank about the limit of that model. A desktop plugin runs inside the app with the app’s own authority and can do anything the app can. Review and commit pinning are the protection, not isolation. Plugins installed from a raw Git URL get a prompt but no review. Treat a plugin as you would any program you install, and prefer catalogue entries.
Skills are lower risk but not risk-free, because a skill is a set of instructions the agent will act on. The Skills pane lists installed skills with on and off switches and a catalogue of optional ones, and the public Skills Hub installs through hermes:// links after you confirm.
MCP servers are configured in settings as well, and v0.23.0 added price tags and spend limits for paid MCP tools.
Privacy: what leaves your machine
Hermes Desktop is local software, but several features contact other services. This is what the documentation describes:
| Feature | What is sent, and where |
|---|---|
| Local model | Nothing after the model download, according to the Local Models page: no account, no API key, no network access |
| Cloud model provider | Prompts, file contents the agent reads, and tool results go to the provider you chose |
| First-run setup chat | Runs on a free trial model, nous/welcome, hosted by Nous Research, once per operating-system user; Skip setup avoids it |
| Nous Portal | Optional sign-in. Portal requests carry a client tag with the Hermes version |
| Update check | Asks the GitHub API for the latest version, using a GitHub token from your environment or the GitHub CLI if one exists |
| Web and browser tools | Fetch the pages the agent decides to visit |
| Theme search | Queries the VS Code Marketplace when you browse themes |
| Remote or Hermes Cloud connection | The conversation runs on the remote host, which is also where commands execute |
The desktop documentation we read does not describe product analytics for the app. That is an absence in the docs, not a guarantee, so a team with strict rules should watch the app’s network traffic before approving it.
For a fully local setup: skip the setup chat, choose a local model as the main model under Settings > Model, check that auxiliary models are not still pinned to a cloud provider (the app warns when they are), and leave web tools off for confidential work.
Safety settings and known limits
An agent that runs shell commands needs guard rails. Hermes checks commands against a list of dangerous patterns, and approvals.mode in config.yaml decides what happens next:
| Mode | Behaviour |
|---|---|
smart (default) | An auxiliary model rates the risk: low-risk commands run, dangerous ones are denied, uncertain ones ask you |
manual | Every flagged command asks you |
off | No prompts at all, the same as YOLO mode |
The status bar has a per-session YOLO switch. Leave it off on a machine that holds credentials. smart mode relies on a model’s judgement, so use manual where a mistake would be costly, or run the terminal back end in Docker. The security guide also lists a hard block list that applies in every mode, protected file paths and scanning of project files for prompt injection.
Limits worth knowing before you standardise on it:
- Platform gaps. No Intel Mac bundle, no packaged Linux release, and the MSIX package needs Windows 11 22H2.
- A fast release train. Five tagged releases appeared between 23 October and 10 November 2026. Good for fixes, harder for change control.
- Unsandboxed plugins. The catalogue’s review is the boundary.
- Local model demands. The 64K context floor rules out small GPUs for anything beyond the smallest models.
- Built for one person. Profiles separate configurations, not people. There is no central policy for which models, plugins or tools a group of users may use.
If you want a desktop agent with a different trade-off, our LM Studio Bionic guide covers a closed-source alternative with a document sandbox.
How VDF AI fits
Hermes Desktop gives one person a capable agent on their own computer, and its safeguards are personal: an approval mode you pick, plugins you choose to trust, a model you configure. VDF AI does not run, host, resell or integrate with Hermes Desktop or Hermes Agent. It is a separate platform for the organisational case, where the questions are who approves models and tools, who may use which data, and where the record of each agent action is kept.
For that case, VDF AI provides governed agents, multi-agent Networks and private RAG on your own servers, air-gapped if needed, or in VDF AI Cloud, with role-based access control and audit logs on every plan. VDF AI Router registers Ollama and other on-premises model deployments next to the cloud models your policy approves, and its air-gap mode keeps routing on local models only.
Sources
Product page, documentation, repository files and release notes as they stood on 14 November 2026.
- Nous Research: Hermes Desktop download page and Hermes Agent site
- Hermes Agent repository: NousResearch/hermes-agent, desktop app README and releases (v0.16.0, v0.22.0, v0.23.0, v0.23.1)
- Hermes documentation source: Hermes Desktop, Local Models, installation, platform support, native Windows
- Hermes documentation source: providers, security, plugins, plugin catalogue
- Ollama: Hermes Desktop integration and context length
- llama.cpp: server README and build b11521
- Community projects: fathah/hermes-desktop and dodo-reach/hermes-desktop
Need agents for a whole organisation rather than one desktop? See VDF AI Agents or book a demo.