Nous Research Announces Hermes Business and Hermes Enterprise

The announcement lays out three pillars for the Business account. Each team member uses agents across multiple channels; the cost comes out of one central balance, with caps that can be set per member; and any skill someone creates is shared with the whole team and accumulates as an asset that belongs to the company. In the post, Nous Research wrote that these shared skills "compound into proprietary IP." Enterprise is the edition that puts the same capabilities on-premises or in the cloud of the customer's choice, and the company called it "a complete, self-improving, sovereign AI stack."

Seen through a marketer's eyes, this announcement is not a product description but the structure of a sales page transcribed word for word. The Business page on Nous Portal splits the features into three verbs: automate, elevate, proliferate. The first is collective wisdom, in which the best skills anyone on the team discovers are automatically shared across the whole organization; the second is spend intelligence, in which one harness across every AI model gives full visibility on cost and return; the third is owning your frontier, in which every skill, memory and workflow the team builds compounds into proprietary capability. The page introduces the product as "a self-improving agent for every team member."

The terms of the Enterprise edition are on the same page. It runs on infrastructure the customer controls, comes with tailored deployment, SSO, service-level agreements and white-glove onboarding, and is built around how the organization already works. In the X post, Nous Research wrote that the stack is already trusted by some of the world's largest companies, but the names of those companies appear nowhere in the announcement.

To understand what changes, you have to look at how this agent grew. METAL has reported that Hermes Agent is built to create its own skills and stack them up every time it finishes a job. That property, preserving memory across sessions, running scheduled tasks and turning finished work into reusable skills, had until now stayed inside individual accounts; the Business account moves that skill library to the team level. A report template or a data-retrieval routine one person built runs as-is inside a colleague's agent at the next desk. It is like donating a single book to a library and finding the same book on the shelf in every reading room the next day.

The numbers come from secondary reporting. According to reports, Hermes Agent passed 175,000 GitHub stars within four months of its February release, and a later snapshot cited by Nous Research put the project at roughly 214,000 stars and nearly 40,000 forks. Nous Portal is the gateway that unifies 248 models plus hosted tools such as web search, image generation and a browser behind one login, and the Business account layers a shared balance and team controls on top of that gateway. METAL reported on that structure last month, when the portal opened an Upstage model free for a week. Before the team tier, the reports said, each developer managed separate API keys, subscriptions, usage limits and memory stores.

The reports also put numbers on why skill sharing turns into money. According to the reports, Nous Research presented internal benchmarks in which agents with more than 20 self-created skills used about 40% fewer tokens and finished similar recurring tasks about 40% sooner than fresh instances. Those figures came from repetitive work such as report generation and data consolidation; one-off, cross-domain tasks performed about the same as fresh instances, and the model configuration and sample size were not disclosed. The saving is large for a team that produces the same report every month, and not so large for a team that gives its agent a different job each time.

The shared balance also acts as a safety device. According to the reports, administrators can cap each member's usage so that one person's task cannot consume the entire team balance. It is a guard against an agent stuck in a loop, calling the same tool over and over, turning into a bill for the whole team. The announcement identified skills as the shared knowledge layer and did not say that conversation histories or personal agent memories are visible to every member.

What Enterprise sells is a data boundary. According to the reports, this edition packages the agent, memory, skill system and model integrations for deployment inside infrastructure the customer controls, and the core can run without telemetry or tracking. Because Hermes Agent is not tied to a particular model, a company can pair it with open-weight models running on its own GPUs or in a private cloud, while requests that go out to hosted models, search APIs or browser providers fall under those vendors' data policies. METAL has reported that Nous Research told users to fork or self-host the agent however they wanted; Enterprise is that freedom wrapped in a contract, SSO and a service-level agreement.

The 25-second launch video METAL reviewed has no spoken words. Between blue-toned skyscrapers and night streets, a hand holds out a business card that reads Hermes Business, and the same name appears at the end. A company that used to address the open-source community chose this time to hand over a business card. Within half a day of posting, the message had passed 119,000 views and drawn 107 replies.

According to reports, Nous Research has been in talks to raise funding at a $1.5 billion valuation. GitHub stars are not revenue, and team subscriptions and enterprise contracts are the shortest route to turning those stars into recurring revenue. The skill library stays with the company when an employee leaves, and the more it fills up, the more it costs to switch to another agent. What Nous Research sold in this announcement is not the agent but that library.