Hermes Desktop: The Open-Source Agent Moves onto the Desktop
Nous Research launches Hermes Desktop, an open-source desktop agent covering macOS/Windows/Linux that reuses the CLI agent's full skills and memory — usable with just mouse clicks.
Hermes Desktop: The Open-Source Agent Moves onto the Desktop
Nous Research launches Hermes Desktop, an open-source desktop agent covering macOS/Windows/Linux that reuses the CLI agent's full skills and memory — usable with just mouse clicks.
Open-source AI agents used to live in the command line. You had to know how to type commands, configure environments, and read error messages. Now Hermes has torn that barrier down. Hermes Desktop wraps the previously CLI-oriented agent into a desktop client with a graphical interface, covering all three major platforms: macOS, Windows, and Linux.
What Is Hermes Desktop
Hermes Desktop is the native desktop app for the Hermes Agent from Nous Research. It isn't a new product — it's a graphical wrapper around the CLI version of the Hermes Agent, reusing the exact same agent, skills, and memory.
In other words, the agent you got comfortable with in the terminal is the same one under a new desktop shell. The capabilities haven't changed; the interaction has just moved from typing commands to clicking a mouse.
On OpenRouter's global daily call-volume leaderboard, Hermes has already overtaken OpenClaw for the top spot. As of this writing, the gap has widened to 660 billion tokens versus 249 billion.
Core Features
GUI + System Tray
Configuration for models, tools, credentials, and more lives in a visual settings panel. You can send your first message within seconds of the first launch.
Streaming Tool Calls
You can watch it call tools step by step as it works. A side-by-side preview on the right renders web pages, files, and tool results while you chat.
Project Management
Sessions can be organized by project, keeping different tasks, contexts, and workflows separate from one another. A built-in file browser lets you leaf through the working directory without leaving the app.
Voice Conversations
Voice interaction is supported: say things to it, and it can talk back to you.
Cross-Entry Memory Sync
The conversation history in the desktop window is the same one from your CLI and your Telegram. Switch entry points and the memory never breaks.
About "Running Locally"
When you see the words "local," "desktop," and "open source," don't rush to conclusions.
Hermes is essentially an agent framework, not a model. What runs on your computer is the agent's runtime environment, skills, and configuration — the "brain" doing the actual thinking still needs to be connected externally. What you connect is up to you:
- Nous Portal (300+ models plus a tool gateway)
- OpenRouter (aggregates 200+ models)
- NVIDIA NIM
- OpenAI
- Your own endpoint
A single hermes model command switches providers — no code changes, no lock-in. Inference latency, privacy boundaries, offline or not: it all depends on which model you pick and how you configure it.
OpenClaw Compatibility
Hermes is compatible with OpenClaw. One migration command, hermes claw migrate, moves OpenClaw's configuration, memory, and skills over in their entirety.
Permissions and Security
A desktop agent can browse your files, invoke the terminal, drive your browser, and store its memory and skills locally. Hermes devotes a security section in its documentation to seven layers of defense in depth: command approval, container isolation, DM pairing, and other mechanisms.
Its skill system can also create skills on its own and improve itself through use. Finish a complex task and it distills new skills out of it. Capabilities are growing, and so are the boundaries of permissions and data access. Before you let it through the door, be clear about what it may and may not touch.
Installation
- macOS / Windows: download the installer directly
- Linux: one-line install
- Windows build: automatically sets up Python, Node.js, ripgrep, ffmpeg, and portable Git, with no administrator privileges required
Download: https://hermesagent.nousresearch.com/desktop
GitHub: https://github.com/NousResearch/hermesagent/tree/main/apps/desktop
Who It's For
- People who want to try an open-source agent but can't use the command line: the graphical interface lowers the barrier to entry
- Existing Hermes CLI users: the same agent and memory, one more entry point
- People who want model choice: no vendor lock-in, switch freely
- Users who care about open source and privacy: MIT license, auditable source code
Related articles
Running LLMs on-Device, Hands-On: Building a Local Agentic Coding Environment with Pi + LM Studio
With the Pi agent framework, an LM Studio inference server, and a Docker sandbox, you can run the Gemma 4 series on a local M2 Mac for agentic coding, linting, and unit tests — reaching roughly 75% of frontier-model accuracy.
WeChat "Xiaowei" in Gray-Scale Testing, Hands-On: 12 Entry Points That Tuck AI into Your Information Feed
A hands-on look at WeChat's native AI assistant "Xiaowei" in gray-scale testing. The main model is WeLM, and 12 entry points cover high-frequency scenarios like chat history retrieval, document summaries, Official Account digests, and local life services.
Baidu DuMate, A Practical Guide: From Installation to Office Automation
A full walkthrough of DuMate, the general-purpose office agent from Baidu, covering installation, skills, app connections, and automation — get up and running in 3 minutes and hand your daily office chores to AI.
DeNovoSWE: The First Long-Horizon Doc2Repo Training Set, Teaching Code Agents to Build Repositories
Renmin University's Gaoling School has released the DeNovoSWE dataset — 4818 real task instances that train Code Agents to generate complete repositories from documentation, lifting Qwen3-30B from 5.8% to 47.2% on BeyondSWE-Doc2Repo.
Doubao Seed 2.1 Pro, Hands-On: Coding Enters the Top Tier, with Multimodal Surprises
A hands-on review of ByteDance's Doubao Seed 2.1 Pro: agent coding and multimodal capability have crossed the production-ready line, including rebuilding front-end interactions from screenshots, at a price nearly 80% lower than Claude Opus 4.6.
Hyper3D Rodin Gen-2.5: A Million Polygons in 4 Seconds as Thinking Comes to 3D Generation
Deemos has released Hyper3D Rodin Gen-2.5, the first to bring an LLM-like Thinking mechanism to 3D generation — million-polygon models in 4 seconds, 10-million-polygon precision, and native 12K texturing.