An open-source AI agent that builds reusable skills from experience, ships 40+ built-in tools, and runs on a $5 VPS — or almost nothing when idle. Nous Research’s Hermes Agent has racked up 8,700 GitHub stars in weeks. But does it actually dethrone OpenClaw?
The AI agent space is getting crowded. Every week brings another framework promising to be your “personal AI assistant.” Most of them are chatbot wrappers with tool calling bolted on — impressive for a demo, underwhelming in practice.
Hermes Agent, released by Nous Research in early 2026, takes a different approach. Its core pitch isn’t just “it can use tools” — it’s “it gets better at using them over time.” After completing a complex task, the agent automatically generates a reusable skill document. Next time it encounters something similar, it loads that skill instead of figuring things out from scratch. Use the skill enough, and it starts refining itself.
It’s an ambitious claim. I spent time digging through the codebase, documentation, and community discussions to see whether it holds up — and how it compares to OpenClaw, the framework that’s been quietly powering personal AI agents since 2025.
What Is Hermes Agent?
Hermes Agent is an MIT-licensed, Python-based AI agent framework built by Nous Research — the lab behind the Hermes, Nomos, and Psyche model families. As of late March 2026, it has 8,700+ GitHub stars, 142 contributors, and 2,293 commits.
Unlike most agent frameworks that tie you to a single model provider, Hermes is provider-agnostic from the ground up. You can use OpenRouter (giving access to 200+ models), OpenAI, Anthropic, z.ai/GLM, Kimi/Moonshot, MiniMax, or your own self-hosted endpoint running Ollama, vLLM, or SGLang. Switching providers is a single command — no code changes required.
The agent runs persistently on your own infrastructure: a $5/month VPS, a Docker container, a remote SSH server, or serverless platforms like Modal and Daytona that hibernate when idle and cost almost nothing between sessions. You interact with it through Telegram, Discord, Slack, WhatsApp, Signal, or a full terminal UI — all from a single gateway process.
The Closed Learning Loop: Where Hermes Genuinely Innovates
Most AI agents have memory in the same way a goldfish has a long-term plan. They maintain context within a session, maybe store some conversation history, and call it a day.
Hermes attempts something more ambitious: a closed learning loop that works across four layers:
Layer 1: Curated Memory
The agent maintains two bounded files — MEMORY.md (~2,200 characters for environment facts, conventions, and lessons) and USER.md (~1,375 characters for user preferences and communication style). These are injected into the system prompt at session start. The agent manages both files autonomously: adding entries when it learns something useful, replacing outdated information, and consolidating when space runs tight.
Layer 2: Full-text Session Search
All past conversations are stored in SQLite with FTS5 full-text search. When the agent needs to recall something from a previous conversation, it searches its history and uses LLM summarization to extract relevant context — on-demand, without bloating the system prompt.
Layer 3: Automatic Skill Creation
After completing a complex task (typically involving 5+ tool calls), the agent can autonomously generate a skill — a structured markdown document with procedures, pitfalls, and verification steps. Next time a similar task comes up, the agent recognizes the pattern and loads the skill directly.
Layer 4: Skill Self-Improvement
Skills aren’t static. When the agent discovers a better approach while using a skill — a faster method, a common edge case, a more reliable tool sequence — it updates the skill document in place.
This loop is the genuine innovation here. OpenClaw has a skill system and a self-improvement skill, but the creation and refinement process is more manual. Hermes automates the entire cycle.
Feature Comparison: Hermes Agent vs. OpenClaw
| Feature | Hermes Agent | OpenClaw |
|---|---|---|
| Language | Python | Node.js/TypeScript |
| License | MIT | MIT |
| Memory System | MEMORY.md + USER.md + SQLite FTS5 | MEMORY.md + memory/ daily logs + vector search |
| Skill Creation | Automatic after complex tasks | Manual install or creation |
| Skill Improvement | Self-updating during use | Available via add-on skill |
| Model Providers | OpenRouter (200+), OpenAI, Anthropic, z.ai, Kimi, MiniMax, custom | OpenAI, Anthropic, Google, custom |
| Terminal Backends | 6 (Local, Docker, SSH, Daytona, Singularity, Modal) | Local + SSH |
| Messaging Platforms | Telegram, Discord, Slack, WhatsApp, Signal | Telegram, Discord, Slack, WhatsApp, Signal, Feishu |
| MCP Support | Native | Via plugin |
| Cron Scheduling | Built-in | Built-in |
| Installation | curl one-liner |
npm install -g |
| Research Tools | Batch trajectory gen, RL environments, trajectory compression | Limited |
| Feishu Integration | None | Full native (docs, bitable, wiki, drive) |
| OpenClaw Migration | hermes claw migrate |
— |
Where Hermes Agent Leads
Self-learning is real, not marketing. The automatic skill creation and self-improvement loop isn’t vaporware — it’s implemented in the codebase and works as described. After using Hermes for a series of SEO audits, it created a skill capturing the entire workflow: tool selection, common pitfalls, output formatting. On the next audit, it loaded the skill and completed the task in roughly half the steps.
Serverless backends change the economics. Daytona and Modal support means your agent environment hibernates when idle and wakes on demand. For users who don’t need 24/7 uptime, this can reduce hosting costs to near-zero during inactive periods.
Research-ready out of the box. If you’re working on training better tool-calling models, Hermes includes batch trajectory generation, Atropos RL environments, and trajectory compression. This positions it as both a user-facing product and research infrastructure.
Community momentum is strong. 8,700 stars in a few weeks is notable. The contributor base (142 people) suggests healthy community engagement beyond a corporate-backed project.
Where OpenClaw Still Wins
Feishu integration is a dealbreaker for Chinese users. OpenClaw has deep, native Feishu support — document read/write, Bitable (multidimensional tables), knowledge base navigation, cloud storage operations, and permission management. For teams and individuals in China’s enterprise ecosystem, this isn’t optional — it’s essential. Hermes has zero Feishu support and no announced plans to add it.
The Node.js ecosystem matters. If your team works in JavaScript/TypeScript, OpenClaw’s architecture is more natural to extend, debug, and contribute to.
Production maturity counts. OpenClaw has been running in production longer. Edge cases around message delivery, error recovery, multi-session management, and cross-platform quirks have been ironed out through real-world usage.
ClawHub skill marketplace. OpenClaw has ClawHub, a centralized skill marketplace where users can discover, install, and share agent skills. While Hermes has a Skills Hub, it’s still in early stages.
Installation and Quick Start
Getting Hermes Agent running takes about two minutes:
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bashWorks on Linux, macOS, WSL2, and Android via Termux. Native Windows is not supported — you’ll need WSL2.
After installation:
source ~/.bashrc # reload shell
hermes setup # interactive setup wizard
hermes # start chatting in the terminal
hermes model # choose your LLM provider
hermes gateway # start messaging gateway (Telegram, Discord, etc.)Coming from OpenClaw? There’s a built-in migration path:
hermes claw migrate # imports settings, memories, skills, and API keysCommon slash commands will feel familiar: /new to start a fresh conversation, /model to switch providers mid-session, /skills to browse available skills, /stop to interrupt the current task.
The Verdict
Hermes Agent is the most interesting agent framework to emerge in 2026. Its closed learning loop — automatic skill creation, self-improvement during use, and bounded persistent memory — represents a genuine step forward in how AI agents can accumulate expertise over time.
But “most interesting” isn’t the same as “best for everyone.”
If you’re a developer or researcher who wants to push the boundaries of what autonomous agents can do — and you don’t depend on Feishu — Hermes Agent is worth your attention. The self-learning capabilities are real, the architecture is clean, and the community is active.
If you’re working in a Feishu-heavy environment, or you value production stability over cutting-edge features, OpenClaw remains the more practical choice. Its Feishu integration alone makes it indispensable for a significant chunk of users.
The good news: both are MIT-licensed and free. There’s nothing stopping you from running both and deciding for yourself.
The AI agent space is better off having both projects competing and cross-pollinating ideas. We all win from that.
- Hermes Agent Documentation: hermes-agent.nousresearch.com/docs
- Hermes Agent GitHub: github.com/NousResearch/hermes-agent
- OpenClaw Documentation: docs.openclaw.ai
- OpenClaw GitHub: github.com/openclaw/openclaw
Comments