Hermes Agent is an open-source, self-evolving AI agent from Nous Research. In May 2026 it topped OpenRouter's global application usage chart with single-day token consumption of 271 billion, ahead of popular tools like OpenClaw and Claude Code. Its core selling point isn't parameter count but "growability" -- the more you use it, the better it knows you.

What Is Hermes

Hermes is an AI Agent with a self-learning loop. Unlike an ordinary chat assistant, it can:

• Automatically create "Skills" from task experience and keep optimizing them in later use
• Hold persistent memory, search past conversation records, and keep context across sessions
• Build a "User Model" that grows to know your preferences and habits as interactions deepen
• Schedule timed tasks and automatically run repetitive work like daily reports and backups

In short: it isn't a chatbot that only answers questions, but an assistant that gets work done, remembers, and learns.

Why It Suddenly Blew Up

On May 9, 2026, Hermes topped OpenRouter's global app token consumption chart. Key numbers:

As of publication, Hermes has consumed more than 6.37 trillion tokens cumulatively; GitHub stars have reached 140,000 (doubled within a month), with 21.8k forks.

The top five most-called models: MiMo-V2-Pro (Xiaomi), MiniMax M2.7, Nemotron 3 Super (NVIDIA), Step 3.5 Flash (StepFun), and Hy3 preview (Tencent). Chinese-made models take the lion's share.

Core Features, Broken Down

Multi-platform access: one Agent covers every scenario

Hermes isn't limited to the terminal. It supports messaging platforms including Telegram, Discord, Slack, WhatsApp, Signal, and Email, and manages every channel through a unified Gateway process. A task started on Telegram can continue in the terminal -- the conversation carries over.

The self-learning loop

This is the capability that separates Hermes from other agents:

1. Automatic skill creation: after finishing a complex task, Hermes automatically extracts the reusable workflow and saves it as a skill
2. Continuous skill optimization: every time a skill runs, it improves automatically based on execution results
3. Cross-session memory: built on FTS5 full-text search + LLM summaries, it can recall past conversations precisely
4. User modeling: using the Honcho dialectic method, it builds an increasingly accurate model of your preferences

Supports nearly every mainstream model

One-command switching via hermes model, no code changes required:

• OpenRouter (200+ models)
• OpenAI, Anthropic
• NVIDIA NIM (Nemotron)
• Xiaomi MiMo, DeepSeek, Zhipu GLM, Kimi/Moonshot, MiniMax
• Any self-hosted endpoint compatible with the OpenAI API

Scheduled tasks and parallel execution

• Built-in cron scheduler defines scheduled tasks in natural language (daily reports, weekly audits, etc.)
• Can spawn isolated sub-agents for parallel workflow processing
• Write Python scripts that invoke tools over RPC, compressing multi-step pipelines into a single call with zero context cost

Flexible deployment options

Seven terminal backends: local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox. Daytona and Modal provide serverless persistence -- asleep when idle, woken on demand, at almost no cost.

Quick Install

Linux / macOS / WSL2

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

After installation, restart the terminal:

source ~/.bashrc    # or source ~/.zshrc
hermes              # start a conversation

Windows (native PowerShell, Early Beta)

irm https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.ps1 | iex

Native Windows support is still in early testing; for the most stable experience, use the Linux install command inside WSL2.

Basic Usage

Common commands after installation:

Common slash commands to use inside a conversation:

• /new or /reset — start a new conversation
• /model [provider:model] — switch models
• /skills — browse the skills you have created
• /compress — compress context to save tokens
• /stop — interrupt the current task

Migrating from OpenClaw

If you were previously using OpenClaw, Hermes offers one-command migration:

hermes claw migrate              # Interactive migration (full migration)
hermes claw migrate --dry-run    # Preview what gets migrated without executing
hermes claw migrate --preset user-data   # Migrate user data only, not keys

What gets migrated: the SOUL.md persona file, memory entries, custom skills, the command whitelist, messaging platform configuration, and API keys.

Compatibility in China

Hermes has officially announced support for China's major platforms:

• Cloud providers: Alibaba Cloud, Tencent Cloud, Huawei Cloud (all offer one-click deployment)
• LLMs: DeepSeek, Zhipu GLM, and Kimi are on the official support list
• Work messaging: Feishu, DingTalk, and WeCom support one-click integration
• WeChat: pair by scanning a QR code and use it directly

Who It Is For

• Developers: those who need a coding assistant that remembers context and executes tasks on its own
• Content creators: use scheduled tasks to auto-generate reports and organize material
• Productivity tool enthusiasts: those who want an AI assistant in Telegram/WeChat that truly "gets you"
• OpenClaw users: migrate seamlessly and gain stronger memory and learning abilities

Things to Note

Hermes's usage numbers are eye-catching, but keep the following in mind:

• Token cost: 271 billion tokens a day shows it can work, but it also means inference isn't cheap. Pair it with cost-effective models
• Stability: as a fast-moving open-source project, some features are still iterating
• Security: for production use, always configure command approval and container isolation

Related Links

• GitHub repository: NousResearch/hermes-agent
• Official docs: hermes-agent.nousresearch.com
• OpenRouter leaderboard: openrouter.ai/app
• License: MIT (fully open source and free)