A first look at Mariwave, from the practice that builds it.
Most AI tools on a Mac can open every file on it. A consultant pastes a client’s revenue model into one to draft the board deck. A studio drops unreleased artwork into another. A yacht broker asks a third to summarize a buyer’s financials before a sea trial. The tool runs with the owner’s own access: the files, the saved passwords, the office network.
At Cisco I led the AI strategy for Webex Contact Center. Our customers were enterprise, federal and regulated organizations, and in those rooms the first question about any AI feature was what it could touch. Years earlier, at Apple, I started Keynote and built the programs that carried iWork into schools and businesses, so the same product had to earn the trust of a teacher and of a company.
In Mariwave, every conversation runs in its own virtual machine that holds only the files you add. It has no network connection of its own, and your passwords and provider keys never go inside it. The model you choose today runs on your Mac or in the cloud, through your own account with a model marketplace at the marketplace’s price, and every tool the AI runs is recorded in a hash-chained log on your Mac.
Consider three desks. A management consultant works from confidential research and ships a deck by Friday. A creative studio turns a brief and a folder of references into images and a presentation. A broker, whether the deal is a portfolio, a building, a painting or a boat, handles one client’s private papers and must later show how each recommendation was made. In March I wrote that the unit of work has migrated from documents to delegated processes. Each of these desks now delegates work to AI on material it has promised to protect.
Arthur Koestler called something that is whole in itself and part of something larger a holon: a cell, an organ, a body; a team, a division, a company. A bank’s contact center and a one-person studio are the same shape at different sizes. One unit of AI work, sealed to what it was given and recorded, repeats at every level. That is the argument of Holon Levels of Agentic LLM and Orchestration, which I later applied to marketing in Holon Levels of Advertising Intelligence. Eyzaguirre & Co. designs that unit for large companies; Mariwave puts it on one Mac.
Some of this appeared on the blog before it appeared in the market. In March 2026, Who Owns the Stack When the Stack Starts to Think? argued that the neutral layer between users and models was “worth too much to leave unclaimed”; about five months later, Stripe agreed to acquire the largest LLM routing marketplace for more than $7 billion. In September 2025, The Philosophy Wars argued for “guided agency through structured permissions”; a month later the leading coding agent shipped a sandbox, and five months later Docker made one generally available. In February 2026, The Jevons Paradox of Security argued that cheaper agents need “explicit constraints”; about eleven weeks later a popular local model server was found leaking memory to anyone who asked (CVE-2026-7482), and five months later a desktop AI assistant’s sandbox was shown reaching the whole Mac. In April, The Bounded Frontier of Agentic ROI called for “constrained, observable, mathematically bounded tools”, five months before Docker took its sandboxes to the cloud.
Early access is open at mariwave.com.