Hermes Agent

Project Overview

Hermes Agent is an open-source personal AI agent built by Nous Research, the AI research lab known for the Hermes series of fine-tuned models.^[1] Billed as "the self-improving AI agent" and "the agent that grows with you," it combines the personal assistant capabilities of projects like OpenClaw with a built-in learning loop and research tooling for generating training data and running reinforcement learning experiments.^[2]

The project's trajectory since its public launch on February 25, 2026 has been extraordinary. It went from 862 stars at the end of February to roughly 99,000 stars in eight weeks — the fastest-growing open-source agent framework of 2026 by Dealroom's count^[3] — and stands at approximately 190,500 stars and 33,000 forks as of June 11, 2026.^[2] What distinguishes Hermes Agent from the broader personal agent ecosystem is its dual identity: it's both a practical daily-driver assistant and a research platform for advancing agentic AI.

What It Does

Hermes Agent serves as a self-hosted AI assistant accessible across multiple messaging platforms, with built-in tooling for AI research workflows.

Core capabilities:

• Multi-channel messaging — Telegram, Discord, Slack, WhatsApp, Signal, and CLI from a single gateway process^[2]
• Native desktop app — macOS, Windows, and Linux app shipped June 2026 with one-click install, in-app self-update, drag-and-drop files, and remote gateway login over OAuth^[4]
• Closed learning loop — Autonomous skill creation after complex tasks, skills that self-improve during use, and FTS5 search over its own past sessions
• Cross-platform continuity — Start a conversation on Telegram, continue it on Discord
• Persistent memory — MEMORY.md, USER.md, and SOUL.md files maintain context across sessions
• Skills system — Compatible with the agentskills.io open standard, with a trusted Skills Hub including NVIDIA's skills tap^[5]
• Cron scheduler — Built-in scheduling with delivery to any connected platform
• Subagent spawning — Parallel workstreams for complex tasks
• Voice memo support — Process voice messages natively
• Full TUI — Terminal UI with multiline editing and slash commands

Research capabilities:

• Batch trajectory generation — Generate tool-calling training data at scale
• Atropos RL environments — Reinforcement learning environments for agentic tasks
• RPC tool access — Agent can write Python scripts that call its own tools programmatically

How It Works

Hermes Agent uses a gateway architecture similar to OpenClaw, with a central process managing all messaging channels and routing to the AI model.

Architecture:

1. Gateway — Python process handling channel connections, session management, and message routing
2. Agent runtime — Model-agnostic execution via Nous Portal, OpenRouter, or custom endpoints
3. Skills — agentskills.io-compatible directories that extend agent capabilities^[5]
4. Terminal backends — 5 options: local, Docker, SSH, Singularity, Modal
5. Memory system — File-based persistent memory (MEMORY.md, USER.md, SOUL.md)
6. Clients — TUI/CLI, web dashboard with a browser-based admin panel (MCP catalog, credentials, webhooks, OIDC login), and the Electron desktop app^[4]

Key design decisions:

• Python-based — Installs via curl one-liner, uses uv package manager
• Model-agnostic — Nous Portal (subscription for zero-config), OpenRouter (200+ models), NVIDIA NIM, NovitaAI, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, or bring your own VLLM/SGLang endpoint^[2]
• DM pairing — User authorization via direct message pairing system
• 5 sandbox backends — More isolation options than most alternatives, from local execution to Modal cloud
• Runs anywhere — Positioned for a $5 VPS, a GPU cluster, or serverless infrastructure, decoupled from your laptop

Development Velocity

Since this profile's original date (February 27, 2026), the project has shipped on a roughly weekly release cadence — v0.9 through v0.16 between April 13 and June 5, 2026.^[2] The June 5 v0.16.0 "Surface Release" alone covered 874 commits, 542 merged PRs, and 170 community contributors, and introduced the native desktop app, the web admin panel, remote-gateway authentication, a fuzzy-searchable model picker, /undo, and a security round (a CVE-2026-48710 Starlette pin, SSRF hardening, and subprocess credential stripping).^[4]

An ecosystem is forming around the core repo: NousResearch's own hermes-paperclip-adapter (1.6K stars), which runs Hermes as a managed employee inside a Paperclip company, drew early attention — though it has had no commits since April 4, 2026 and looks dormant relative to the main project.^[6]

Business Model

Hermes Agent is free and open source under the MIT license.^[2]

Nous Research monetizes through their Portal subscription service (v0.16 added a "Quick Setup via Nous Portal" onboarding path) and their broader AI research and model training business, not through the agent software itself.^[4]

Strengths

• Explosive, verifiable momentum — ~190K stars in under four months, growing roughly 9,500 stars/week versus OpenClaw's ~3,000 as of spring 2026^[3]
• Research lab pedigree — Built by Nous Research, a respected AI research organization with deep model expertise^[1]
• Self-improving by design — The only major personal agent with a closed learning loop: autonomous skill creation, in-use skill refinement, and cross-session recall
• Dual-purpose — Practical personal agent AND research platform for training data generation
• 5 sandbox backends — Local, Docker, SSH, Singularity, Modal — more options than any competitor
• Atropos RL integration — Unique reinforcement learning capability for agentic task training
• agentskills.io compatible — Open standard for skills interoperability^[5]
• Mainstream-ready packaging — Native desktop app removes the terminal barrier that limits most self-hosted agents^[4]
• Weekly release cadence — 170 community contributors in a single release cycle

Weaknesses / Risks

• Issue backlog ballooning — Roughly 20,000 open issues as of June 2026; hypergrowth is outpacing triage, and a community-filed security posture review of the repo is among them^[2]
• Security surface grows with capability — An agent with shell access, messaging credentials, and self-modifying skills is a real attack surface; the project itself shipped a CVE pin and SSRF hardening in v0.16, and experienced users advise running it sandboxed^[7]
• Python dependency — Heavier runtime than Go/Rust alternatives; not ideal for edge/embedded
• Fewer channels than OpenClaw — Six messaging surfaces (Signal was added since launch) versus OpenClaw's 50+; still no iMessage or Google Chat
• Research-oriented complexity — Atropos/batch features add complexity that pure personal-use cases don't need
• Nous Research dependency — Project direction tied to a single research lab's priorities
• Ecosystem unevenness — Satellite repos like hermes-paperclip-adapter have gone quiet since April 2026 while the core sprints ahead^[6]

What Developers Say

Hacker News reception has been notably positive, with the caveats you'd expect for a self-hosted agent with broad system access:

"I've been using the NousResearch Hermes agent for the past couple of weeks and have to say it's been really good." — rnxrx, who uses it across multiple machines for Obsidian organization, research, home automation troubleshooting, and server management, noting it improved through its self-reinforcement loop^[8]

"I'm using Hermes. The same applies to all agents, don't give it free reign over all your stuff. Run it within a sandbox." — Flere-Imsaho, in a thread about an agent privilege-escalation vulnerability^[7]

Other HN threads from spring 2026 express surprise that people are "still using OpenClaw" given Hermes Agent's momentum — a sentiment shift worth weighing against OpenClaw's much larger installed base.

Competitive Landscape

vs. OpenClaw The story flipped in 2026. OpenClaw still has the larger ecosystem and 50+ channels, but Hermes Agent's ~190K stars put the two in the same league, and Hermes was growing roughly three times faster as of spring 2026.^[3] Hermes offers 5 sandbox backends (vs 2), a native desktop app, research tooling (Atropos RL, batch trajectories), and Python over TypeScript. Choose Hermes for the learning loop and research capabilities; choose OpenClaw for maximum channel coverage.

vs. NanoBot NanoBot (22K stars at last check) focuses on simplicity with pip-install setup and strong China IM support. Hermes Agent adds research tooling, the self-improvement loop, and more sandbox options, and now has a far larger community.

vs. PicoClaw / ZeroClaw PicoClaw (Go) and ZeroClaw (Rust) optimize for minimal resource usage. Hermes Agent is Python-based and significantly heavier, but offers research capabilities neither has.

vs. HermitClaw Both target research use cases, but HermitClaw is autonomous research (no prompting needed) while Hermes Agent is a full personal assistant with research tooling built in.

Ideal User

• AI researchers who want a personal agent that also generates training data
• Nous Research model users who want tight integration with Nous Portal
• Python developers who prefer Python over TypeScript/Node ecosystems
• ML engineers exploring reinforcement learning for agentic tasks
• Self-hosters who need multiple sandbox backends (especially Singularity for HPC or Modal for cloud)
• Non-terminal users — the June 2026 desktop app makes it viable for people who would never touch a TUI^[4]

Bottom Line

Hermes Agent went from niche curiosity to category heavyweight in under four months. At its February 2026 launch it was an 862-star research-lab side project; by June 2026 it is a ~190K-star phenomenon with a weekly release cadence, a native desktop app, an MIT license, and the fastest growth curve of any open-source agent framework this year.^[3] Its unique pitch — the only personal agent built by a major AI research lab, with a closed self-improvement loop and RL tooling — is now backed by a community large enough to sustain it.

The honest caveats: a ~20K open-issue backlog, a security surface that demands sandboxing (the maintainers and the community both say so), and satellite ecosystem repos that haven't kept pace with the core. If you want maximum messaging-channel coverage, OpenClaw still wins on breadth. For nearly everyone else in the personal-agent space — and especially anyone who cares about an agent that genuinely improves with use — Hermes Agent has become the default choice rather than the research-flavored alternative.

---