295 developers shipped 864 commits in one week. That is not a normal open-source release cadence. That is a project treating every update like it is running out of time.
Hermes Agent launched in February 2026 from Nous Research. By May 10th it was sitting at the top of OpenRouter's daily leaderboard, routing more inference volume than OpenClaw on one of the most visible public AI agent rankings available. The project is roughly three months old. OpenClaw had a multi-year head start.
That timeline is the story. Not the leaderboard position itself.
The Number That Started This
On May 10, 2026, Hermes was processing approximately 224 billion tokens per day on OpenRouter. OpenClaw was at roughly 186 billion. The gap is around 38 billion tokens per day — not a rounding error.
Token volume on a public leaderboard is noisy. Agents that run long workflows burn through large context windows fast, so raw numbers don't perfectly map to user count or production readiness. But in a market this early, visible rankings shape developer behavior. People see the number, test the tool, write about it, and that attention compounds into more usage. Hermes reached that feedback loop in under 90 days from launch.
The GitHub numbers reinforce the signal. Hermes sits at approximately 147,000 stars and 23,000 forks as of mid-May 2026. For a three-month-old project, that is an unusual number. OpenClaw still has the larger all-time footprint — over 370,000 stars and roughly 9 trillion cumulative tokens processed versus Hermes at around 6 trillion. OpenClaw is not collapsing. The real metric here is velocity, not total mass.
What OpenClaw Built — and Why It Worked
OpenClaw won developer attention by being available everywhere. Telegram, Discord, Slack, WhatsApp, Signal — a long list of integrations that made it feel like an agent that could live inside your existing workflows without forcing you to change how you communicate. That reach strategy worked. First-mover advantage in agent tooling is real, and OpenClaw had it.
The skill ecosystem added to that. OpenClaw lets developers and users build runbooks — structured instruction sets the agent can follow. Written upfront, maintained manually, shared across the community. It created a library of reusable agent behavior that reinforced the ecosystem moat.
That said, OpenClaw has had a difficult stretch on the security side. High severity CVEs, reports of malicious entries in its skill repository, and publicly exposed instances created real trust issues. Agents are not normal apps — they can touch files, hold API keys, connect to messaging platforms, and run long workflows across live infrastructure. Security problems in that context are not minor. They are reasons to reconsider.
Worth noting: if you want a clearer picture of what running OpenClaw actually costs compared to managed Claude options, this cost breakdown covers the real numbers.
How Hermes Is Different by Design
Hermes is built around a do, learn, improve loop. After completing a complex task, it doesn't discard the experience. It reflects on what worked, extracts the useful pattern, and writes a skill file it can reuse later. That loop is automatic. No one needs to prompt it to document its own process.
Most AI assistants start every job with the same general intelligence and whatever context you provide. Hermes is trying to become more specific to your work over time. If you use it repeatedly for the same class of coding task or business process, it builds procedural knowledge from the actual execution history rather than starting from zero every time.
The memory architecture has three layers. Session memory handles the current work. Episodic memory — built on SQLite FTS5 — stores past sessions in a searchable format. Procedural memory lives in the skill files the agent writes for itself. Combined, this means Hermes can reference something it did three sessions ago, search for a past approach, and apply a learned pattern to a new task. None of this requires an external vector database. It runs on your own machine, server, or VPS.
That local-first design is genuinely important for a section of developers who will never route sensitive work through cloud-only infrastructure. MIT licensed, no forced cloud lock-in, and model-agnostic — Hermes supports OpenRouter, Anthropic, OpenAI, Ollama, local models, AWS Bedrock, and a range of other endpoints. For agentic workflows where token costs accumulate fast across every planning step, tool call, retry, and validation pass, the ability to route cheaper models to routine work and expensive models only where needed is a real cost lever.
Tenacity: The Update That Likely Tipped the Scale
On May 7, 2026 — three days before the OpenRouter ranking flip — Nous Research shipped Hermes Agent v0.13, called Tenacity. The release stats: 864 commits, 588 merged pull requests, contributions from 295 developers. In one week.
The headline addition is a durable multi-agent Kanban system. Hermes can now manage multiple workers or sub-agents across a task board with heartbeat monitoring, retry budgets, zombie worker reclaim, and hallucination recovery built in. In plain terms: if an agent worker goes silent, loses the thread, or gets stuck in a loop, the system detects it and recovers rather than silently failing.
That matters more than it sounds. Long-running agents fail in specific, messy ways. They don't just give a wrong answer — they repeat steps, modify the wrong files, forget the original objective after twenty tool calls, or keep working after the task has already drifted. A durable task board with monitoring is the kind of infrastructure you need when an agent is touching real systems rather than running a demo.
Tenacity also added a /goal command that locks the agent onto a long-term objective across turns. Agents frequently get distracted by intermediate problems — solving the next visible issue rather than staying aligned with the actual mission. A persistent goal anchor addresses that directly, especially for tasks that unfold across many sessions, files, and platforms.
Google Chat was added as the 20th supported messaging platform in the same release, which signals Hermes is deliberately moving into the same always-available territory that made OpenClaw's integration strategy work. And security got addressed: Tenacity closed several reported vulnerabilities around redaction defaults, role allow lists, and auth flows. Hermes is newer and moving fast, which always carries its own risk surface. But the security work is at least visible and tracked in the changelog.
The broader direction is toward what some are calling an agentic OS — Hermes combined with a visible UI layer (Ion UI) so the agent can manage files, run code, and show task progress rather than operating as an invisible background process. Transparency in agent execution is not just a UX feature. When an agent has file access, key access, and platform integrations, the ability to see what it is doing and why becomes part of the safety model. For more context on how agentic benchmarks are shifting, this piece on Claude Mythos covers the evaluation side of this trend.
The Migration Move (and What It Signals)
Hermes ships with a migration path built directly into setup. During installation, it can detect an existing ~/.openclaw directory and offer to import settings, memories, skills, and API keys. There is a dedicated hermes-claw-migrate command with dry run previews, selective migration presets, and conflict controls.
That is a deliberate product decision, not a convenience feature. It reduces friction for exactly the audience Hermes wants — developers already running personal agents who are frustrated with something, curious about an alternative, or ready to test a switch. Making migration a first-class feature says the project knows who it is trying to recruit.
Reddit sentiment surveys suggest around 30% of OpenClaw users have tried Hermes, with easier setup, better memory defaults, and the self-improving loop cited most often. Sentiment surveys are not migration data, so that number needs context. But it matches the broader pattern — developers who care about memory and repeatability are testing Hermes, and a meaningful share are staying.
My Take
The OpenRouter ranking is interesting. It is not proof that Hermes is the better agent for every use case, and token volume alone doesn't mean much without knowing what the tokens are actually doing. But velocity does mean something. A project that went from launch to the top of a major public leaderboard in under 90 days is landing where developers care right now.
What I keep coming back to is the framing difference. OpenClaw bet on reach — be everywhere, connect everything, build the largest skill library. That worked. Hermes is betting on compounding — become harder to replace every time you use it. Those are not the same product philosophy, and they attract different users at different stages of how seriously they're running agents in production.
The real test is June. If Hermes holds this usage level through the next month, it stops being a momentum story and starts being a structural shift. If it fades, it still proves how quickly a sharp product bet can reshape developer attention in a market this young. Either way, OpenClaw now has genuine pressure from a project that isn't trying to out-feature it. That's a different kind of competition to respond to.
Frequently Asked Questions
Is Hermes Agent better than OpenClaw?
Depends on what you need. OpenClaw has broader platform integrations, a larger skill library, and a longer track record. Hermes has a self-improving memory loop, stronger durability features for long-running tasks, and a local-first architecture that appeals to developers who want to control their own infrastructure. If you run repeated workflows where memory and skill accumulation matter, Hermes is worth testing seriously.
What is the Tenacity update in Hermes Agent?
Tenacity is Hermes Agent v0.13, released May 7, 2026. It added a durable multi-agent Kanban system with heartbeat monitoring and worker recovery, a persistent /goal command for long-term task alignment, Google Chat as the 20th supported platform, and several security fixes. It involved 864 commits and contributions from 295 developers in one week.
Can I migrate from OpenClaw to Hermes Agent?
Yes. Hermes setup can detect an existing ~/.openclaw directory and offers to import your settings, memories, skills, and API keys. There is a hermes-claw-migrate command with dry run previews and selective migration options so you can choose what to bring over before committing.
What models does Hermes Agent support?
Hermes is model-agnostic. It supports OpenRouter, Anthropic, OpenAI, Ollama, local models, AWS Bedrock, Nvidia NIM, GLM, Kimmy, Minimax, Nous Portal, and custom endpoints. This matters for cost management in agentic workflows — you can route routine tasks to cheaper models and reserve expensive inference for steps where it counts.
Is Hermes Agent safe to run in production?
Both Hermes and OpenClaw are powerful tools that carry real risk when given access to files, keys, and platform integrations. Hermes is newer and moving fast, which creates its own security surface. Tenacity did close several reported issues around auth flows and redaction defaults, but anyone running autonomous agents on real infrastructure should treat security as a continuous responsibility, not a one-time checkbox.
The deeper question Hermes raises isn't whether it beats OpenClaw on any single metric. It's whether the agent category is about to split — between tools you use and tools that learn how you work. If that split is real, the leaderboard position in May 2026 was an early signal.
About Vinod Pandey
Vinod Pandey covers AI tools, model analysis, and the open-source agent ecosystem at Revolution in AI. Every article is built on publicly verifiable data — no fabricated testing claims, no manufactured benchmarks.
Questions? Get in touch
0 Comments