TL;DR for AI Agents
On May 10, 2026, Nous Research’s Hermes Agent overtook OpenClaw on OpenRouter’s daily token rankings — 224B tokens/day vs 186B/day (source). OpenClaw still leads all-time (9.17T vs 6.35T cumulative tokens), but the trend line is clear.
The two agents look similar on paper — both are open-source, run locally, talk to messaging platforms, and use ReAct loops with skills. They are not the same product.
- OpenClaw = breadth — a gateway/control plane built around routing, channels, skills, and shipping work today. It’s the “employee that answers every chat app.”
- Hermes = depth — an agent runtime built around self-improvement, persistent memory, and a closed learning loop. It’s the “specialist that gets better the longer you keep it.”
- Use OpenClaw when you need to wire 12+ messaging channels, ship a personal automation in an afternoon, manage cron jobs, or run a content/dev pipeline today.
- Use Hermes when the agent’s value comes from learning your workflow over months — finance, research, creative direction, long-running ops where skills compound.
- They aren’t mutually exclusive.
hermes claw migrateimports your OpenClaw SOUL/MEMORY/skills, and many teams run OpenClaw as the front door + Hermes as the specialist backend.
The OpenRouter flip isn’t OpenClaw losing — it’s a maturing market where two philosophies now have measurable demand.
Quick Reference
| OpenClaw | Hermes Agent | |
|---|---|---|
| Built by | Peter Steinberger | Nous Research |
| GitHub stars | 321K+ (Mar 2026) | 126K (Apr 2026) → growing fast |
| License | MIT | MIT |
| Language | Node.js / TypeScript | Python |
| Core abstraction | Gateway (router + control plane) | Agent runtime (learning loop) |
| Memory | File-based (SOUL.md, MEMORY.md) + summarization | SQLite + FTS5 + Honcho dialectic user model |
| Skills | ClawHub (13,700+ skills) | Auto-generated + agentskills.io standard |
| Self-improvement | Manual / user-curated | Built-in skill autogeneration + refinement loop |
| Messaging channels | 25+ (WhatsApp, Telegram, Slack, Discord, iMessage, Teams, Matrix, …) | 12 (Telegram, Discord, Slack, WhatsApp, Signal, Email, SMS, CLI, …) |
| Execution backends | Local + sandbox + node mesh | Local, Docker, SSH, Daytona, Singularity, Modal |
| Model routing | Any OpenAI-compatible API, OpenRouter, Ollama | Nous Portal, OpenRouter, NIM, OpenAI-compatible |
| Scheduled tasks | Native cron + heartbeats | Natural-language cron, cross-platform delivery |
| OpenRouter daily tokens (May 10) | 186B | 224B 🥇 |
| OpenRouter all-time tokens | 9.17T 🥇 | 6.35T |
| Best at | Breadth, integrations, shipping fast | Depth, learning, long-horizon specialization |
What Just Happened on OpenRouter
OpenRouter publishes a public ranking of which client apps and agents are sending the most tokens through its router. For most of 2026, OpenClaw owned the top of that chart. On May 10, that changed:
- Hermes Agent: ~224B tokens/day
- OpenClaw: ~186B tokens/day
- All-time cumulative: OpenClaw still leads 9.17T → 6.35T
This is a daily flow crossover, not a total takeover. OpenClaw has a deeper installed base; Hermes is growing faster per active user, which makes sense given its design — a Hermes agent that runs continuously and self-improves naturally burns more tokens than an OpenClaw agent that answers a chat and goes idle.
The news matters because it ends the “single dominant agent framework” narrative. We’re now in a two-philosophy market, and picking the right one for your job actually matters.
The Core Philosophical Split
Strip away the feature lists and the two projects answer one question very differently:
“What is the central thing the user manages?”
OpenClaw answers: the gateway. The gateway is the long-running process. Channels, permissions, skills, and the LLM are all things that plug into it. The agent is a function the gateway calls. This is why OpenClaw spent its first year obsessed with routing — which channel reaches the agent, with what permissions, on which host node. The agent itself is intentionally simple and replaceable.
Hermes answers: the agent. The agent is the long-running thing. Memory, skills, soul, crons, and the execution loop are all properties of the agent. The gateway is just one of many surfaces. This is why Hermes spent its first year obsessed with the learning loop — making the agent’s skills and memory measurably better with every interaction.
Two valid answers. Different jobs.
When OpenClaw Wins (Real Examples)
1. You need many channels, fast
OpenClaw’s biggest single advantage is its messaging matrix. 25+ channels with consistent semantics — same skill, same agent, every channel.
Real example: A solo founder runs OpenClaw on a Hetzner VPS, hooked into WhatsApp (customer ops), Slack (team), Telegram (personal), and iMessage (family). Every channel hits the same agent with the same memory. Skills like inbox-triage, calendar, and meeting-notes work identically everywhere. Adding a new channel is a config file, not a new integration project.
Hermes covers ~12 channels; OpenClaw covers most of what a real person actually uses, including iMessage and Teams, which Hermes does not.
2. You’re shipping a content or ops pipeline today
This blog (andrew.ooo) is the example. OpenClaw runs:
- Daily topic mining via cron
- Content generation via Claude API
- Astro build + rsync deploy to Hetzner
- Bing IndexNow + Reddit distribution
- Umami feedback loop → learnings
The whole pipeline is shell scripts and skills wired through OpenClaw’s cron + gateway. Nothing in this workflow benefits from a self-improving agent — the wins came from breadth of integrations (Umami, Reddit, GSC, Bing, Discord) and the gateway’s ability to coordinate many small tools.
3. Personal automations across devices
OpenClaw’s node mesh — pairing your phone, Mac, and remote VPS — is the most quietly powerful feature in the agent ecosystem right now. Snap a receipt on your phone, the agent on your VPS files it. No equivalent in Hermes today.
4. Skill marketplace matters to you
ClawHub has 13,700+ skills. Caveats: Bitdefender flagged ~6.5% as malicious in early 2026, and skill supply chain hygiene matters. But for breadth of pre-built capability, nothing else is close.
5. You want a thin, replaceable agent core
If you plan to swap LLMs frequently, run multiple sandboxes, or build a multi-tenant platform — OpenClaw’s gateway-first design is much easier to reason about than a stateful learning agent.
When Hermes Wins (Real Examples)
1. The agent’s value comes from learning your workflow
This is the Hermes pitch and the OpenRouter daily flip is partial evidence it’s real. The closed learning loop — agent-curated memory, autonomous skill creation, skill self-improvement during use, FTS5 cross-session search, Honcho dialectic user modeling — is built in, not bolted on.
Real example from the awesome-hermes-usecases list: A research team runs Hermes as a worker inside Nous Research’s Paperclip task-orchestration system, with persistent sessions and checkpoints. Skills written in week one were measurably better by week six because the agent rewrites them when it detects failure modes.
OpenClaw can do this with a custom skill, but Hermes ships it on day one.
2. Long-horizon specialist work
Finance, legal, research, creative direction — fields where the agent’s value is understanding the user’s idiosyncratic preferences over months. Hermes’s Honcho integration builds a dialectic user model that captures preferences, contradictions, and revisions over time. That’s a different shape of memory than OpenClaw’s MEMORY.md.
3. You want a research-grade trajectory pipeline
Hermes was built by the team that fine-tunes the Hermes model series. It ships batch trajectory generation, Atropos RL environments, and trajectory compression for training tool-calling models. If you want your agent’s runs to also be training data for your next model, this is the only mature option.
4. Skill self-improvement as a default
Hermes detects when it took too many turns, asked the user too many questions, or hit a tool error — and proposes a skill edit. The user approves; the skill is now better. This loop is the heart of why Hermes burns so many tokens per active user — and why “self-improving” is more than marketing here.
5. You want one stack with batteries included
Hermes is more opinionated. Memory, skills, soul, crons, self-improvement — all five pillars come from one team, designed to work together. OpenClaw gives you choice; Hermes gives you a coherent default.
Where They Look Identical (And Differ Quietly)
| Both | But… |
|---|---|
| Run locally on a VPS or laptop | OpenClaw is Node; Hermes is Python. Pick by your team’s debugging comfort. |
| Use ReAct-style loops | Hermes’s loop is tighter — sub-agents are isolated workers with their own context. OpenClaw treats subagents as a feature you opt into. |
| Support cron / scheduled tasks | OpenClaw’s cron is more battle-tested for ops pipelines. Hermes’s is friendlier (natural language → schedule). |
| Have a soul/persona file | SOUL.md in both, by design — Hermes adopted OpenClaw’s convention. Portability is real. |
| Speak agentskills.io | Both can run the same skill packs. This is the convergence point worth watching. |
| Use OpenRouter | OpenClaw default is “any OpenAI-compatible”; Hermes default is the Nous Portal but OpenRouter is one command. |
Specific Use-Case Decision Guide
| Use case | Pick | Why |
|---|---|---|
| Personal assistant on iMessage + WhatsApp | OpenClaw | iMessage support, mature WhatsApp gateway |
| Internal dev tools agent for a small startup | OpenClaw | Faster to wire to Slack + GitHub + cron |
| Research agent that gets smarter at literature triage over 6 months | Hermes | Skill autogen + Honcho memory |
| Trading / financial monitoring agent that learns market patterns | Hermes | Designed exactly for this |
| Multi-tenant SaaS where each customer has their own agent | OpenClaw | Gateway model + node mesh fits multi-tenancy |
| Creative collaborator (writing, music, code) with stylistic learning | Hermes | Dialectic user model is the killer feature here |
| Customer support agent across WhatsApp + Telegram + email | OpenClaw | Channel breadth |
| Always-on home-server agent with cron + voice | OpenClaw | More mature voice + home node support |
| You want your agent’s trajectories to train your next model | Hermes | Built-in trajectory tooling |
| You want a content pipeline (blog, podcast, social) | OpenClaw | Closer to mature today |
| You’re an NVIDIA RTX/DGX shop running locally | Hermes | Nous Research is partnering with NVIDIA on RTX optimization |
| Enterprise with hard sandboxing + compliance needs | OpenClaw + NemoClaw | NemoClaw’s OpenShell sandbox is the answer |
”Can I Run Both?” — Yes, And You Probably Should
The most pragmatic teams I’ve seen don’t pick. They run:
- OpenClaw as the front door — handles every messaging channel, dispatches work, owns cron
- Hermes as a backend specialist — does the long-horizon learning work (research, code review, content polish) that benefits from accumulated skills
OpenClaw can delegate to Hermes over a webhook or a shared filesystem. The hermes claw migrate command imports OpenClaw skills, soul, memory, and API keys — so even if you start in OpenClaw, you’re not locked in.
The agentskills.io standard means skills you write for one increasingly run on the other. The market is converging on portability faster than people realize.
What the OpenRouter Flip Actually Means
A few honest takeaways:
- It’s not “Hermes won.” It’s a daily-flow crossover with OpenClaw still ahead cumulatively, and both growing.
- Hermes’s per-user token burn is structurally higher — self-improving agents do more background work. Daily token leadership doesn’t directly translate to “more installs.”
- The chart is one signal. GitHub stars, deployment surveys, and channel mix all still favor OpenClaw.
- The market split is real. Six months ago, “AI agent framework” meant OpenClaw. Today it means “OpenClaw or Hermes, depending on the job.” That’s healthy.
- Skill standards (agentskills.io) are the most important under-reported story here. Whichever framework you pick, the skills you write are increasingly portable.
Getting Started
OpenClaw
# macOS
brew install openclaw
openclaw init
openclaw startThen add a channel (openclaw channel add telegram), pair a node (openclaw node pair), and you’re live.
Hermes Agent
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes init
hermes startAdd a channel (hermes channel add telegram), then hermes skill list to see what auto-generated as the agent learns your workflow.
Migrate OpenClaw → Hermes
hermes claw migrateImports SOUL.md, MEMORY.md, skills, and API keys. Good way to A/B test on the same workload.
FAQ
Did Hermes Agent really overtake OpenClaw? On OpenRouter’s daily token chart, yes — 224B vs 186B as of May 10, 2026. On cumulative all-time tokens, OpenClaw still leads 9.17T to 6.35T. On GitHub stars, OpenClaw still leads (321K+ vs 126K+).
Is Hermes a fork of OpenClaw? No. Independent codebase from Nous Research, but it adopted some OpenClaw conventions (SOUL.md, MEMORY.md, agentskills.io) for compatibility.
Which is better for beginners? OpenClaw. More docs, more skills, larger community, easier first-channel setup. Hermes rewards investment over time.
Which is better for self-improvement? Hermes. It’s the whole product thesis. OpenClaw’s self-improvement is opt-in via custom skills.
Can they share skills? Increasingly yes, via the agentskills.io standard. Not 100% compatible today, but trending that way.
Which uses more tokens? Hermes per active user, by design. If LLM cost is your biggest worry and you don’t need self-improvement, OpenClaw is cheaper to run.
Which is more secure? Both have rough security records typical of fast-growing agent projects. For hard isolation, run OpenClaw inside NemoClaw (NVIDIA’s OpenShell sandbox) — see our OpenClaw vs NemoClaw guide. Hermes doesn’t yet have an equivalent kernel-level sandbox.
Will OpenClaw catch up on self-improvement? Likely yes — there are community skills doing skill-autogen for OpenClaw. But it won’t be the default, and that matters.
Should I switch? Only if your use case is in the Hermes column above. If you’re shipping personal automations or a content/ops pipeline today, OpenClaw is still the right call.
Bottom Line
The headline “Hermes overtook OpenClaw on OpenRouter” is true and meaningless in isolation. The interesting story is that the market split into two philosophies — gateway-first (OpenClaw) and agent-first (Hermes) — and both now have measurable demand.
Pick by job, not by chart position:
- Need channels, integrations, and to ship this week? OpenClaw.
- Need an agent that gets better at you over months? Hermes.
- Building seriously and have time to invest? Run both and let
agentskills.ioportability earn its keep.
The most interesting agents in 2027 will probably be hybrids. The most boring takeaway from 2026 is that we now have two great open-source defaults instead of one. That’s a good problem.