Somewhere inside a large company, someone is moving a number from one window to another. The applications work. The employee knows the procedure. Yet the transaction needs a person to carry it across the gap. H Company has built its business around that gap: the stubborn distance between software that stores information and software that can finish a job.
- H builds AI that operates existing software through screens, code and tools.
- Its current range runs from a free Chrome companion to enterprise agents deployed with engineers.
- The buying question: can it finish your workflow reliably, at an acceptable total cost?
The Paris company’s proposition is pleasingly practical. A computer-use agent reads an interface, decides where to click, types, and checks what happened. That matters when the application has no convenient API, the programmed doorway that lets software exchange instructions. Businesses have plenty of those applications. Their employees supply the missing connections, one copy-and-paste at a time.
A very expensive beginning
H’s origins date to late 2023, with French incorporation following in April 2024. Charles Kantor joined forces with four former DeepMind researchers: Laurent Sifre, Karl Tuyls, Daan Wierstra and Julien Perolat. Their expertise included reinforcement learning, game theory and multi-agent systems. Sifre had worked on projects including AlphaGo and Gemini. This was a team assembled to build models, with the scientific credentials to attract attention.
In May 2024, H announced $220 million in initial financing. Accel, Amazon, UiPath, Bpifrance and prominent individual investors were among the backers. UiPath offered an especially interesting connection: an established automation business investing in a company proposing another way to automate. Accel described the relationship as a potential distribution channel for H’s models.
The first publicly visible rupture came before a product release. In August, Tuyls, Wierstra and Perolat departed. H cited operational differences. By June 2025, Gautier Cloix, a Palantir veteran, had replaced Kantor as CEO. A formidable founding roster had proved insufficient to keep the founding team together.
The engineer comes with the agent
Cloix brought experience with forward-deployed engineers: people who work directly with customers to make software function in the customer’s actual environment. Speaking in July 2025, he described an initial ambition to concentrate on two or three large customers. The useful unit of progress was becoming a working deployment.
That approach now appears throughout H’s offering. Its research lab trains models; product development packages them; deployment engineers confront the customer’s systems. H describes a sequence of discovery, prototyping, an agreed proposal, production launch and expansion. The sequence contains a lesson buyers can copy: establish the workflow and its success measures before expanding the assignment.
“Ship fast, learn faster.”H Company’s published team values
The company reports banking deployments that assemble onboarding documents, query registries and flag risks for analysts. Its energy examples include gathering scattered billing information and preparing invoices, plus reconciling purchase orders and receipts for human sign-off. H reports a 90% reduction in time-to-invoice in one example. These are company-reported outcomes from unnamed deployments; they describe use cases rather than establish a general productivity guarantee.
One company, several entrances
In June 2025, H introduced a suite with memorable names: Runner H coordinated workflows, Surfer H navigated the web, and Tester H translated instructions into visual software tests. The company also released Holo1 model weights and WebClick, a dataset containing 1,639 interaction scenarios. It was selling products while giving builders material to inspect and reuse.
The current lineup makes the layers clearer. HoloTab is a free Chrome side panel for everyday browsing tasks. Users can type or speak instructions, record a routine, and schedule it. H advertises email drafts for approval, travel research and marketplace listings. Developers can instead use the Models API for inference or the Agents API for managed execution. Enterprises can buy a deployment built around their own processes.
Holo4, released in September 2026, combines screen interaction, coding and tool calls. A workflow can change methods as it moves between applications. That flexibility is central to H’s position among computer-use models, agent platforms and conventional automation. Its enterprise claim rests on controlling the model, orchestration and deployment together, including work on legacy systems.
Read the small receipt
H lists Agents API plans at $0, $20 and $200 a month, plus custom enterprise terms. Session allowances are approximate because tasks differ in complexity. The Models API charges by tokens. A free browser extension and a negotiated enterprise deployment plainly buy different things.
The release’s downloadable weights also deserve a careful look. Holo4’s 35B-A3B version uses Apache 2.0; the 27B version carries a Creative Commons noncommercial license. “Open weights” tells you something about access. The individual license tells you what you may do with that access.
Long tasks remain revealing. In October’s research notes on post-training an NVIDIA Nemotron model, H reported that supervised fine-tuning improved shorter benchmarks while lowering its OSWorld 2.1 score from 28.4 to 24.9. The team responded by merging the trained checkpoint with the base model and restoring selected attention weights. The merged result reached 39.4. More training had required a change of recipe.
That experiment supplies a useful rule for customers: test the whole task. An agent that finds every button may still lose the thread across twenty steps. Automation becomes a poor fit when errors are costly, supervision consumes the savings, or the workflow exceeds the agent’s demonstrated reliability. H’s own enterprise design retains human decisions at consequential points.
The thing to borrow is a small, measurable trial: one process, explicit permissions, an inspectable action record, and an agreed finish line. The number copied between windows is a modest starting point. It is also a better test of H’s promise than another impressive conversation with a chatbot.