On May 10, 2026, a quiet milestone reshaped the AI agent landscape. Hermes Agent by Nous Research overtook OpenClaw to claim the #1 spot on OpenRouter's global daily token rankings — processing 224 billion tokens in a single day. Just three months after its February 2026 launch, Hermes had accumulated over 140,000 GitHub stars and nearly 1,000 contributors.
Something about this agent is different. That difference is self-evolution.
What Is Hermes Agent?
Hermes Agent is an open-source, self-improving AI agent built by Nous Research and released under the MIT license. It is not a chatbot wrapper or a prompt template. It is an active orchestration layer — a framework that handles persistent memory, tool use, and skill creation in a continuous loop across sessions.
The core design philosophy: every task is a learning opportunity. When Hermes solves a problem, it does not just return an answer. It writes a skill file capturing the exact steps, edge cases, and domain knowledge it applied. On the next similar task, it retrieves that skill and uses it as working context.
The result is an agent that compounds value over time. The more you use it, the better it gets at the things you actually do.
The Self-Learning Architecture
Hermes follows a three-phase loop:
Do — Execute the task using available tools and memory.
Learn — After tasks requiring five or more tool calls, generate a Markdown skill file summarizing the approach, edge cases, and domain knowledge.
Improve — Store the skill in SQLite with FTS5 full-text search; retrieve and apply it on future similar tasks.
This is not a marketing claim. The learning loop is a concrete, inspectable system. SQLite stores skills with full-text indexing, so retrieval is fast and deterministic. The skills are plain Markdown files — you can read, edit, and version-control them like any other code.
Three-Layer Memory
Hermes splits memory into three distinct layers:
Layer 1 — User Snapshot: A persistent profile of your preferences, work style, and long-term goals. Updated continuously across sessions.
Layer 2 — Session History: Every conversation stored in SQLite FTS5 with semantic search. The agent can recall what it told you six weeks ago, what approach it took on a similar task, or what errors it encountered.
Layer 3 — Skill Files: Procedural knowledge captured as Markdown files after complex tasks. 118 built-in skills ship with the framework; 647 more are available in the community ecosystem.
This architecture means Hermes does not suffer from context window amnesia. It knows what it has learned, and it knows where to find it.
Getting Started
Hermes ships with a single-command installer that handles all prerequisites automatically on Linux, macOS, and WSL2:
curl -fsSL https://hermes-agent.org/install.sh | bashOnce installed, launch the agent:
hermesIt connects to your local model server by default — LM Studio, Ollama, or llama.cpp. For cloud models, set your API key:
export ANTHROPIC_API_KEY=your_key_here
hermes --model claude-sonnet-4-6Hermes works with virtually any model. Claude Sonnet 4.6, Qwen 3.6, MiMo-V2-Pro, and dozens of others are all officially supported. The framework is provider-agnostic by design.
Recommended Models
For local deployment, Nous Research and NVIDIA recommend the Qwen 3.6 series:
- Qwen 3.6-35B runs on roughly 20 GB of VRAM while exceeding 120B model performance in agent benchmarks
- Qwen 3.6-27B matches 400B model accuracy at 1/16th the size — ideal for RTX workstations and NVIDIA DGX Spark
The 35B model is the sweet spot for most developers: excellent reasoning, manageable VRAM requirements, and fast enough for interactive use on modern hardware.
For cloud deployments, Claude Sonnet 4.6 and Gemini 3.5 Flash are both well-tested in production Hermes deployments.
v0.13.0 "Tenacity" — What Changed
Released May 7, 2026, version 0.13.0 "Tenacity" was built by 295 contributors across 864 commits. Key additions include:
- Kanban boards: Visual task management built directly into the agent interface
- Persistent goal-tracking: Goals survive session restarts and load automatically into context
- 8 security fixes: Addressed CVEs in the subprocess execution layer and skill file parser
- 20+ messaging integrations: WhatsApp, Telegram, Discord, Slack, and more through a unified gateway
- Sub-agent isolation: Contained sub-agents now use focused context windows, reducing hallucinations from context overflow
Why Hermes Topped OpenClaw
The OpenRouter rankings comparison tells the story clearly. On May 10, 2026:
| Agent | Daily Tokens |
|---|---|
| Hermes Agent | 224B |
| OpenClaw | 186B |
| Kilo Code | 149B |
| Claude Code | 79.2B |
OpenClaw still leads in cumulative tokens (9.17 trillion vs Hermes's 6.35 trillion) and total GitHub stars (370,000+). But Hermes's daily velocity signals where new development is happening. Developers are choosing Hermes for new workflows, and the gap is widening.
The xAI/Grok endorsement was a turning point: xAI publicly named Hermes Agent as the recommended open-source agent framework for Grok users. NVIDIA followed with a dedicated blog post on running Hermes across RTX hardware and DGX Spark. Over 110,000 GitHub stars arrived in the first three months.
The Business Case for Teams
For development teams evaluating Hermes against proprietary alternatives:
Data stays local: Skills, memory, and session history never leave your infrastructure. For enterprise use cases involving proprietary code, client data, or regulated industries, local deployment is not optional.
No per-seat licensing: MIT license means you can deploy, fork, and modify without vendor negotiation. The cost floor is your own compute.
Compound value: The longer your team runs Hermes, the more specialized it becomes to your stack, conventions, and domain. That accumulated skill library belongs to you.
Active development: 864 commits in a single release, 295 contributors, nearly 1,000 total. The development velocity is among the fastest in open-source AI right now.
Conclusion
Hermes Agent reaching #1 on OpenRouter is more than a usage statistic — it is a signal that developers are ready for agents that genuinely learn. The architecture is transparent, the license is permissive, and the community velocity is exceptional.
If you have been waiting for an open-source agent worth deploying in production, Hermes is worth serious evaluation. Start with a local model, let it run for a week, and watch what it learns.
The GitHub repository is at github.com/NousResearch/hermes-agent. The v0.13.0 release notes are thorough, the documentation is maintained, and the community responds quickly.