Hermes Agent Hit 246,000 Stars. This Week It Started Selling to Enterprises.

Hermes Agent sat at roughly 27,000 GitHub stars in April 2026. It is at 246,184 today, with 51,461 forks. Nous Research launched it in February as the successor to OpenClaw, built around a learning loop that writes its own skills from experience rather than waiting for you to install them. This week the project grew a commercial layer, packaging Hermes for teams plus private infrastructure. The timing is what makes it interesting. Anthropic cut Claude Code weekly limits by 17% on September 14, Codex users spent the same week posting about exhausted quotas, then three separate Hermes ecosystem launches landed inside 24 hours. Hermes runs on whatever model you point it at, which is the entire pitch. Best for developers who want model choice plus self-hosting. Not ideal for anyone who needs procurement-grade guarantees today.

In April, Hermes Agent had about 27,000 GitHub stars.

It has 246,184 now.

That is 219,000 stars in five months for a project most people outside agent tooling still have not heard of, from a lab that was doing open-weight model research three years ago.

This week it did something more interesting than grow. It started charging.

What Hermes Actually Is

Facts below come from the repository plus Nous Research’s own documentation, checked on September 16.

Two numbers in that table deserve a second look, plus one of them is not flattering.

The Learning Loop Is the Whole Pitch

Most agents are stateless. You open a session, you explain your project again, you close the session, the context evaporates.

What Hermes does differently

Hermes creates skills from experience rather than waiting for you to install them. It improves those skills during use, then builds a persistent model of you across sessions.

That is a different product category from a coding assistant. It is closer to a long-lived personal operator that happens to code, which is why it runs unattended plus reaches you across messaging platforms instead of living in a terminal.

Where it came from

Hermes is the direct successor to OpenClaw, which we covered in our OpenClaw deep dive. Anyone who ran OpenClaw will recognise the shape, though the learning loop is the piece OpenClaw never had.

The skills angle

Hermes skills follow the agentskills.io open standard, so they are portable plus shareable through a community Skills Hub rather than locked to the runtime.

That choice matters more than it looks. We traced how agent skills became a vendor channel and found licences splitting in two directions. Hermes picked the open side, which is consistent with shipping the whole runtime under MIT.

Twenty Platforms Is the Unusual Part

Most agents live in one place. Hermes lives in twenty.

The list

CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, email, SMS, DingTalk, Feishu, WeCom, Weixin, QQ Bot, Yuanbao, BlueBubbles, Home Assistant, Microsoft Teams plus Google Chat. All from one gateway, documented here.

The Chinese platform coverage is worth noticing. DingTalk, Feishu, WeCom plus Weixin are not platforms a Western agent project usually prioritises, which tells you where a meaningful share of the adoption is coming from.

Why it changes the use case

An agent you talk to in Telegram at 11pm is a different thing from an agent that lives in your IDE. It handles the tasks you think of away from your desk, plus it delivers results where you already are.

We wrote about Claude Code reaching into Telegram and Discord when that arrived. Hermes started there rather than adding it later, which shows in how the whole thing is arranged.

About Those 43,824 Open Issues

This is the number nobody quoting the star count mentions.

What it might mean

Forty-three thousand open issues is enormous by any standard. On a repo with 246,000 stars it works out around one issue per six stars, which is a high ratio.

There are two readings. The generous one says heavy real usage generates heavy real feedback, plus a project shipping this fast accumulates issues faster than it closes them. The unkind one says triage has collapsed.

What the release cadence says

The cadence argues for the generous reading. Between late March plus mid April the project shipped seven major versions. The two weeks before v0.9.0 saw 209 pull requests merged alongside 81 issues closed. A more recent window landed around 250 commits across 461 files with roughly 125 merged pull requests.

That is not a stalled project. It is a project moving faster than its issue tracker.

Why you should still check

If you are evaluating Hermes for anything that matters, read the issues before the README. Filter for your platform plus your model provider. Projects this large have failure modes that only show up in the tracker, plus the ratio here means there are plenty to read.

What Shipped Recently

Release notes are dull reading, though they answer the question a star count cannot: is anyone still building this.

Pantheon, August 31

Version 0.21.0 brought manual subagent control. You can list live children, steer one mid-flight, then stop another without losing the partial work it already did.

That sounds minor until you have watched a delegated agent spend twenty minutes going the wrong direction. Your only option was killing everything. Delegation is where the tokens go, so correcting a subagent rather than restarting it saves real money.

A patch followed on September 7, which is the cadence you want to see after a major release.

The window before that

An earlier release window landed roughly 250 commits across about 461 files, adding 42,613 lines while removing 1,641, through around 125 merged pull requests.

That work included migration to the MCP 2.x SDK, support for the stateless protocol, a bundled Bot Mode plugin with a teammate protocol, plus runtime hardening around the Python subprocess environment.

Reading the cadence

Seven major versions between late March plus mid April. Two hundred and nine pull requests merged in the fortnight before v0.9.0. This is a project moving at a speed that breaks things, which is worth knowing before you put it near anything load-bearing.

Fast shipping is a feature while you are experimenting. It becomes a liability the moment you depend on stable behaviour, plus the open issue count suggests plenty of people have found that edge already.

The Commercial Turn

Two days ago Nous Research packaged Hermes for teams plus private infrastructure. Team subscriptions along with enterprise contracts.

Why now

The open-source agent has given Nous a large developer audience. The lab reportedly held funding talks at a $1.5 billion valuation in July, raising at least $75 million led by Robot Ventures with participation from USV.

Converting that audience into recurring revenue is the obvious next step. The pitch targets organisations that want model choice, deployment control plus access to agent internals, which is precisely what the closed alternatives do not offer.

The gap in the announcement

Self-hosting the runtime does not mean self-hosting everything. Requests sent to hosted models, search APIs or browser providers still leave the private environment under those vendors’ data policies.

That distinction will matter to anyone buying this for compliance reasons. Running Hermes on your own infrastructure while pointing it at a hosted frontier model means your data still travels.

Procurement teams will also want single sign-on, identity provider integration, audit logs, retention controls, backup procedures, upgrade policies plus support terms with a service level agreement. None of that appeared in the launch announcement.

The Week It Landed In

Here is where the timing gets hard to ignore, though it needs stating carefully.

What else happened

Anthropic ended a temporary 50% Claude Code usage boost on September 14, replacing it with a permanent 25% increase over the older baseline. Net result for anyone using it the week before: about 17% less capacity. Users on $200 plans posted cancellations.

Codex users spent the same week posting about exhausted quotas plus swapping in cheaper models to stretch them.

Then on September 16, three Hermes things launched inside 24 hours. A plugins catalog inside Hermes Desktop. Then iHermes, a personal assistant running on Hermes with iMessage as the interface. Then a writeup from a company describing how their whole team moved onto Hermes, covering what they tried first plus what it cost.

What I am not claiming

I cannot show you that Hermes growth is being driven by people leaving Claude Code. Star velocity on a repo this size is not something the GitHub API exposes cleanly, plus three ecosystem launches in a day is the kind of thing that happens when a team coordinates a release rather than when a market shifts.

What is fair to say is narrower. The open agent whose entire pitch is that it runs on any model you choose grew a commercial tier in the same week both major closed agents gave their users less. Those facts sit next to each other. Draw your own line.

The structural point

Model choice is the hedge, plus it is the only one available. If your agent runs on whatever endpoint you point it at, a vendor tightening limits is an inconvenience rather than a crisis. You change a config value.

Our Claude Pro review covers what the subscription buys at full price. The thing it cannot buy is insulation from the provider changing the deal, which is the argument Hermes makes without having to say it.

Harness Engineering, Briefly

Hermes sits inside an argument that took over agent discourse earlier this year, plus the argument explains why a project like this can matter more than a model release.

The experiment

The LangChain team held the model constant then adjusted only the surrounding harness, meaning the instructions, constraints, feedback loops, memory plus orchestration. Their internal benchmark score moved from 52.8% to 66.5%. Ranking went from outside the top 30 into the top 5.

Zero model changes. Mitchell Hashimoto, who created Terraform, named the discipline Harness Engineering.

Why that matters here

If the harness carries that much of the performance, then the harness is the product. A well-built harness on a mid-tier model can beat a bare frontier model, which reframes what you are actually buying when you pay for an agent subscription.

Hermes is a harness with a learning loop attached, shipped under MIT, pointed at whichever model you like. That is a direct bet on the harness mattering more than the weights.

The same logic runs through the Claude Code repos people actually run daily. Most of the value in those is configuration rather than code.

What Nous Research Is, Since Most People Skip That Part

The lab matters here, because the same organisation trains models plus ships the agent that runs them.

The background

Nous Research was founded in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium plus Shivani Mitra. It built its reputation on open-weight model releases, the Hermes model family among them, long before shipping an agent.

That history explains a design choice that otherwise looks strange. A lab that trains models could have built an agent locked to its own weights. Instead Hermes runs on Nous Portal, OpenRouter, OpenAI or any endpoint you give it.

Why that choice is defensible

Locking the agent to one model would have capped adoption at whoever liked that model. Leaving it open let the agent grow on whichever frontier model developers already pay for. That includes models from direct competitors.

Two hundred and forty-six thousand stars suggests the trade worked. The lab now has distribution that its model releases alone never produced, plus a commercial tier to sell into it.

One more thing to add

The commercial tier arriving two days ago is the first time Nous has asked anyone for money directly. Everything before it was free code plus free weights, funded by investors rather than users.

The tension worth watching

A company raising at a reported $1.5 billion valuation eventually needs the open runtime to feed something that makes money. Right now that is team plus enterprise contracts, which is the cleanest version of this arrangement.

The question is what happens if those contracts do not grow fast enough. The MIT licence protects the code that exists today. It does not promise the next version arrives on the same terms.

The Skills Hub Changes the Maths

One more piece deserves attention, since it connects to something we have tracked for weeks.

Portable by design

Hermes skills follow an open standard rather than a proprietary format, so a skill written for Hermes is not stranded there. The Skills Hub exists for sharing them, plus community contribution is the expected route rather than an afterthought.

Why that matters against the alternative

We counted what shipped in the vendor collections. Unity put out 31 skills under a restricted licence that GitHub cannot classify. Google Maps put out one under Apache 2.0. Neither is portable in the sense that matters, because both are written for a specific product.

An agent accepting community skills in a shared format gets a library it never had to write. Package managers ended up mattering more than the languages around them for the same reason.

The catch

Open contribution means unreviewed contribution. Install a community skill and you are running instructions a stranger wrote, inside a tool holding file access plus terminal access.

We flagged this in the vendor collections too, where 28 of Unity’s 31 official skills declared no tool scoping at all. The problem does not get smaller when the authors are strangers instead of a corporation.

What You Should Actually Do

If you are curious rather than committed

Install it on something disposable. Any cheap VPS works, plus the project ships an install script. Point it at a model you already pay for then give it a task you would otherwise do yourself.

The learning loop needs time before it shows anything. A single session tells you nothing about the thing that makes this different.

If you are shopping after the limit cut

Understand what you are switching to. Hermes is a harness, not a model. You still pay for inference somewhere, so the saving comes from routing cheaper models at the steps that do not need a frontier one.

That is the workaround Claude Code users converged on anyway, which is splitting planning from execution across models with different costs. Hermes makes it the default rather than a trick.

If you are evaluating for a team

Read the issue tracker first, filtered to your stack. Then ask Nous directly about single sign-on, audit logs plus retention, because the launch material does not cover them.

Also map which parts stay private. Self-hosting the runtime while calling a hosted model means your data still leaves, so decide whether that satisfies whoever is asking.

If you already run Claude Code or Codex

Keep them. Hermes is not a replacement for a coding agent sitting in your editor, plus pretending otherwise wastes a weekend.

The realistic setup runs both. Coding agent in the IDE for the work you supervise. Hermes on a cheap box for the long jobs you want running while you sleep.

If you are staying where you are

That is defensible. 43,824 open issues is a real number, plus a project shipping seven major versions in three weeks is a project that will break things.

Our earlier piece on Hermes covers the basics if you want the shorter version before committing an afternoon.

The Part Worth Keeping

Twenty-seven thousand stars in April. Two hundred and forty-six thousand now.

Five months. Somewhere in there Hermes stopped being a project people starred then became a thing people run, which is a transition most open-source agents never make. The forks say that more plainly than the stars do. 51,461 people cloned it to do something.

Now it has a price. Not for the runtime, which stays MIT, but for the layer around it that teams need. Single sign-on, audit trails, someone to call. That is the standard path, plus it usually ends with the free version quietly mattering less.

Watch the issue tracker rather than the star count. 43,824 open issues is either the sound of a project being used hard or the sound of one losing control of itself. Which one it turns out to be will decide this faster than any funding round.

Charts and Blocks

Star growth, April to September 2026

Open harness against closed subscription