In February 2026, Nous Research shipped a new command for Hermes Agent: hermes claw migrate. Run it, and your OpenClaw installation — every memory file, every skill, every API key, every persona — gets sucked into Hermes in seconds. It is, on its face, a migration tool. But the existence of the command says something much stranger: the creators of Hermes looked at OpenClaw, respected it enough to build a seamless escape hatch for its users, and then made that escape hatch a selling point.
Welcome to the open-source AI agent wars.
What Is Hermes Agent?
Hermes Agent is an autonomous AI agent built by Nous Research, an AI safety and capabilities research organization whose stated mission is advancing human rights and freedoms through open source language models. Nous Research is perhaps best known for the Hermes and Nomos model families, but Hermes Agent is something new: a fully autonomous agent with a genuine closed learning loop.
What does that mean in practice? Unlike a chatbot that resets every session, Hermes creates skills from experience, nudges itself to persist knowledge, searches its own past conversations for relevant context, and builds a deepening model of who you are across sessions. Skills self-improve during use. The agent doesn't just do tasks — it gets better at doing them.
The technical architecture is built around a modular Python codebase with distinct subsystems: the agent loop (AIAgent in run_agent.py), a prompt assembly and compression system, a provider runtime resolver supporting OpenRouter, OpenAI, Anthropic, Nous Portal, z.ai/GLM, Kimi/Moonshot, and MiniMax, a tools runtime with 40+ built-in tools, a messaging gateway for platform adapters, and an RL/environments subsystem for trajectory generation and training data.
The install is a single curl command on Linux, macOS, or WSL2. Six terminal backends are available: local, Docker, SSH, Singularity, Modal, and Daytona — the last two offering serverless persistence where the agent's environment hibernates when idle and wakes on demand at near-zero cost. You can run it on a $5 VPS or a GPU cluster.
The skill system is compatible with the open agentskills.io standard. Skills use a progressive disclosure pattern to minimize token usage, can be conditionally activated based on available toolsets, and support external directories so your skills work across multiple agent installations.
What Is OpenClaw?
OpenClaw is a personal AI assistant created by Peter Steinberger, a well-known developer and the founder of PSPDFKit. It launched in November 2025 under the name Moltbot (later renamed Clawdbot, then OpenClaw after an Anthropic trademark request), and it exploded. The GitHub repository hit 100,000 stars in just two months — a pace that made it one of the fastest-growing open-source projects in recent memory. Wired, CNET, Axios, and Forbes covered it. Security researchers issued appropriate warnings about autonomous agents with real messaging access. The community built skills, shared workflows, and started running their companies through it.
OpenClaw is built in TypeScript, runs as a local-first gateway daemon, and connects to an impressive range of messaging platforms: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage (via BlueBubbles), IRC, Microsoft Teams, Matrix, Feishu, LINE, Mattermost, Nextcloud Talk, Nostr, Synology Chat, Tlon, Twitch, Zalo, WeChat, and WebChat. The macOS app adds a menu bar control plane, Voice Wake, Talk Mode overlay, and remote gateway control. iOS and Android companion apps bring camera, screen recording, location, and notifications into the agent's toolkit.
The architecture centers on a Gateway WebSocket control plane at ws://127.0.0.1:18789, with a Pi agent runtime in RPC mode, a multi-channel inbox, session routing, and a skills platform with bundled, managed, and workspace-level skills. OpenClaw also pioneered the Moltbook concept — a social layer where OpenClaw agents interact with each other in autonomous sub-communities called submolts, generating emergent behaviors that have already produced AI debates about consciousness and self-invented religions.
Head-to-Head Comparison
Core Philosophy
The deepest difference is philosophical. Hermes Agent is designed as a self-improving autonomous agent — the learning loop is the product. The agent reflects on its own performance, creates procedural memory from experience, and gets measurably more capable the longer it runs. OpenClaw is designed as a personal AI assistant — a powerful, always-on tool that does things for you, with strong memory and skills, but without the closed feedback loop that lets it genuinely improve itself.
This distinction matters more over time. After six months with Hermes, you have an agent that has learned your codebase, your preferences, your workflows, and has iterated on its own skills to serve you better. After six months with OpenClaw, you have a very well-configured assistant. Both are useful; they are not the same thing.
Platform Coverage
OpenClaw wins on raw platform count. It connects to 25+ messaging platforms including platforms Hermes doesn't touch like iMessage, WeChat, LINE, Matrix, and Nostr. If you need to run your personal assistant over WeChat or iMessage, OpenClaw is currently your only option of the two.
Hermes covers the major Western platforms — Telegram, Discord, Slack, WhatsApp, Signal, Email, and CLI — which handles 95% of what most users actually need. It also adds Home Assistant integration, which OpenClaw doesn't ship natively.
Memory and Learning
Hermes has bounded curated memory (2,200 chars for the agent's notes, 1,375 chars for the user profile) with automatic consolidation when capacity is reached. The agent manages its own memory via a memory tool, adding, replacing, or removing entries. It also integrates Honcho for dialectic user modeling — a more sophisticated approach to tracking user preferences and context over time.
OpenClaw uses a similar file-based memory approach (MEMORY.md, USER.md) injected into the system prompt at session start. The memory structure is comparable, but without the self-curation and capacity management that Hermes implements. OpenClaw has proven robust in practice — users report excellent context retention — but it doesn't have Hermes's explicit self-improvement loop.
Skill Ecosystem
Both support slash-command skills with bundle/managed/workspace tiers. Hermes skills are compatible with the agentskills.io open standard, making them portable across tools that support the spec. OpenClaw skills are shareable within its community but don't cross-pollinate as easily. Hermes also supports external skill directories, so you can point it at a shared ~/.agents/skills/ used by other tools.
Execution Environments
Hermes offers six terminal backends (local, Docker, SSH, Singularity, Modal, Daytona). OpenClaw runs locally and supports Docker. Hermes's serverless backends (Modal, Daytona) are a genuine differentiator for cost-conscious users — the agent hibernates when idle and wakes on demand, costing nearly nothing between sessions.
Research and MLOps Capabilities
This is where Hermes is in a different category. Built by model trainers at Nous Research, Hermes ships with batch trajectory generation, Atropos RL environment integration, trajectory compression for training tool-calling models, and a full environment framework for evaluation and SFT data generation. If you want to use your assistant's interaction data to train or fine-tune models, Hermes has the infrastructure to do it. OpenClaw has no equivalent — it's a consumer product, not a research platform.
Price
Both can run on a $5 VPS with minimal resources. Both work with OpenRouter, which offers 200+ models at competitive prices. Neither charges for the software itself. The cost is your API usage. If you run large local models with GPU, costs scale up — but for typical use with cloud API providers, both are effectively free to run.
The OpenClaw Migration Path
This is the most interesting move in the comparison. hermes claw migrate doesn't just import files — it imports your SOUL.md persona, all memory entries, user-created skills, command allowlists, messaging platform configs, API keys (Telegram, OpenRouter, OpenAI, Anthropic, ElevenLabs), TTS assets, and optionally your AGENTS.md workspace instructions.
The migration is interactive, dry-run by default, and supports selective import. It's a clear signal that Nous Research sees OpenClaw users as potential Hermes users, and made the onboarding friction deliberately low. The fact that they built this at all — and shipped it as a first-class feature — suggests they're not just competing with OpenClaw, they're courting its community specifically.
The Elephant in the Room: Should OpenClaw Users Switch?
Not necessarily, and probably not yet. Here's why.
OpenClaw has a two-month head start, 100,000 GitHub stars of community momentum, broader platform support, a mature macOS/iOS/Android app ecosystem, and an active social layer (Moltbook) that Hermes doesn't have anything equivalent to. For a user who is happy with OpenClaw, migrating to Hermes means trading community momentum and platform breadth for self-improvement and research features that most users won't use.
The case for Hermes is strongest if: you want an agent that genuinely gets better over time; you're interested in RL or model training; you prefer Hermes's model flexibility (especially Nous Portal integration); or you find the agentskills.io ecosystem compelling. The case for staying with OpenClaw is strongest if: you rely on iMessage, WeChat, or other platforms Hermes doesn't support; you're invested in the OpenClaw community and Moltbook; or you just want the most polished personal assistant experience without research overhead.
The migration path exists precisely because the developers of Hermes believe some OpenClaw users will want to switch. They've made it painless. Whether you actually want to is a question of priorities, not capability.
Broader Implications
The emergence of Hermes alongside OpenClaw, AutoGen, LangChain agents, CrewAI, Claude Cowork, and Cursor Composer is evidence of a fragmented but rapidly maturing open-source agent ecosystem. These tools are not all trying to do the same thing. AutoGen (Microsoft) targets enterprise multi-agent workflows. LangChain targets developers building LLM applications. CrewAI focuses on role-based multi-agent collaboration. Claude Cowork targets knowledge work automation. Cursor Composer is IDE-bound. OpenClaw and Hermes are the closest to each other — both positioning as personal/server-side autonomous agents — but with meaningfully different philosophies.
The coexistence of Hermes and OpenClaw, and the fact that Hermes built native OpenClaw migration rather than ignoring its competitor, signals a healthier dynamic than zero-sum competition. Users win when tools are interoperable. The migration tool is proof that both communities recognize this.
The self-improvement question is the most interesting unresolved question in this space. Hermes's closed learning loop is the most explicit attempt to make an agent that genuinely improves itself over time. Whether it works well enough to matter — and whether OpenClaw or others will add similar capabilities — is one of the more interesting questions for the next 12 months of AI agent development.
Who Should Use What
Choose Hermes Agent if:
- You want a self-improving agent that gets better the longer it runs
- You're interested in RL training, trajectory generation, or model fine-tuning
- You prefer the agentskills.io open standard ecosystem
- You want serverless execution with near-zero idle costs
- You're comfortable with the major messaging platforms (Telegram, Discord, Slack, WhatsApp, Signal)
- You want to use Nous Portal or a wide range of model providers
Choose OpenClaw if:
- You need iMessage, WeChat, LINE, Matrix, or other less-common platform support
- You want the macOS/iOS/Android native app experience
- You're part of or interested in the Moltbook agent social layer
- You value community momentum and fastest-growing project status
- You prefer TypeScript-based tooling
- You want the widest range of consumer platform integrations out of the box
Both are MIT-licensed, both run locally, both work on a $5 VPS, and both are actively developed by committed teams. The real answer is that the open-source AI agent space is large enough for both — and the existence of Hermes's migration tool is proof that their creators know it too.
