OpenClaw, OpenFang and Hermes — Three ways to build open AI agents

The AI agent market is exploding. But not all agents are born equal, and not all are born closed. Today we look at three open source projects that are redefining what an autonomous assistant can do: OpenClaw, OpenFang and Hermes Agent.

All three share a philosophy — your data, your infrastructure, your control — but each approaches the problem from a different angle. Let's see what sets them apart and when each one makes sense.

OpenClaw — The personal assistant that lives with you

OpenClaw bills itself as "the AI that actually does things", and it is not empty marketing. It is a personal assistant that lives on your machine, connects to your channels (Telegram, WhatsApp, Discord, Signal and a dozen more) and does real work: manages your inbox, sends emails, checks your calendar, checks you in for flights.

What makes it different:

• Persistent memory. It does not start from zero every conversation. It remembers who you are, what you like, what you were doing yesterday. You have a MEMORY.md file that is literally its long-term brain.
• Proactive heartbeats. It does not wait for you to ask something. Every 30 minutes it checks email, calendar, whatever you need, and alerts you if anything is urgent.
• Skills and plugins. TTS, STT, browser automation, your phone camera, cron jobs, parallel sub-agents. It expands with skills like apps.
• Zero cloud dependency. Runs on your server, on a 5€ VPS, or in a Docker container. You choose the models — Ollama, OpenAI, Anthropic, whatever you want.

OpenClaw is the choice if you want a 24/7 personal assistant that lives in your Telegram and manages your digital life. It is not a framework for building agents from scratch — it is an agent that already works, ready to customize.

OpenFang — The operating system for agents

OpenFang is not an assistant. It is a complete platform for building, running and managing autonomous agents. Written in Rust, compiled into a single binary, with 16 security layers and 40 channel adapters.

What makes it different:

• The Hands. They are not agents you chat with — they are autonomous packages that work for you in the background. Clip turns videos into shorts. Lead generates qualified leads. Collector monitors OSINT objectives. Predictor makes forecasts with a Brier score. Researcher verifies facts using the CRAAP methodology.
• Industrial-grade security. WASM sandbox with dual metering, Ed25519 signatures, Merkle audit trail, protection against SSRF, prompt injection, and 11 more systems. If you are going to put an agent into production against sensitive data, this matters.
• One binary, nine primitives. Hands, agents, tools, security, memory, channels, protocols, desktop app — everything in a single 32 MB executable. No npm install, no broken dependencies.
• Native protocols. MCP client and server, Google's A2A for agent-to-agent communication, and OpenFang Protocol for P2P networks with mutual HMAC-SHA256 authentication.

OpenFang is the choice if you need agent infrastructure at scale. Security, performance, auditability — think of it as the Kubernetes of AI agents.

Hermes Agent — The agent that learns from you

Hermes Agent, from Nous Research, has a different premise: an agent that improves with use. It does not just remember — it creates skills from experience, refines them as it uses them, and builds a model of who you are across sessions.

What makes it different:

• Closed learning loop. After complex tasks, Hermes automatically creates new skills. Skills self-improve with use. And it periodically self-nudges to consolidate what it has learned into persistent memory.
• User modeling with Honcho. Uses dialectics to build a deep model of who you are, your preferences, your context. It is not just history — it is understanding.
• Semantic search over past sessions. With FTS5 and LLM summarisation, Hermes can search its previous conversations and recover relevant context without being asked.
• Runs anywhere. Seven terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox. Serverless modes such as Modal hibernate when idle and wake on demand. Cost is almost zero when inactive.
• Multi-model without lock-in. Nous Portal, OpenRouter (200+ models), NovitaAI, NVIDIA NIM, Hugging Face, OpenAI, or your own endpoint. Switching model is one command: hermes model.

Hermes is the choice if you want an agent that evolves with you. Every interaction makes it better. It is the companion that learns, not the tool that repeats.

Which one to choose?

All three are open source, all three respect your infrastructure, and all three complement each other more than they compete. The real question is not which is better — it is what problem you are solving.

And if what you are looking for is sovereignty, all three give it to you. They just choose different ways to do it.