Use AI without giving every workflow to a third party.

What “private AI” actually means

Most AI products are private in the marketing sense: a vendor promises to handle your data well. Private AI in the operating sense is different. It means you can point at each step of a workflow and say where it executed and what it transmitted. A chat window that sends every prompt to a hosted model, a “private” tenant on someone else’s cloud, and a model running on the laptop in front of you are three different boundaries, even if the interface looks the same.

The useful mental model is destinations. Every feature either keeps data on the device, sends it to infrastructure you operate, or sends it to a third party under that party’s terms. A private setup is one where the default destination is the device, exceptions are deliberate, and policy can enforce those exceptions for a team.

Where data goes in a typical AI workflow

The chat window is only the visible part. A single question can touch model inference, document retrieval, embeddings, web search, a connector, crash reporting, and account sync. The table lists each hop and what Glyph does with it by default.

How Glyph keeps the boundary visible

The model picker is the control. Local models are labelled as local and run through the runtime that ships with the app — Metal on Apple silicon, CUDA, Vulkan, or CPU on Windows and Linux. Cloud models appear only after you add your own provider key, which is stored in the operating system keychain, and every response from a cloud model is attributed to that provider so a user cannot mistake a hosted answer for a local one.

Features that must leave the device say so at the point of use. Web search, the embedded browser, GlyphNet device pairing, and connectors are each enabled deliberately. That is slower to set up than a product that silently does everything in the cloud, and it is the point: the default is on-device, and every exception is a decision someone made.

The same design carries to mobile. Glyph for iPhone and Android runs phone-sized GGUF models on the device and, when you pair it with your desktop over GlyphNet, sends prompts to your own computer rather than to a hosted model.

Private AI for one person versus a team

An individual can download Glyph, pick a model that fits the machine, and have a private workflow in an afternoon. The remaining risks are ordinary device risks: full-disk encryption, backups, and not pasting secrets into a cloud model by accident.

A team has a different problem. Ten people with ten laptops will make ten different choices about cloud models, connectors, and where documents live. Private AI for a team therefore needs policy: which model destinations are allowed, who can enable web access, how a departing employee’s access is revoked, and what evidence exists afterwards. Glyph Private moves those decisions to a server the organisation operates — identity through OIDC, per-user model catalogues, a gateway that routes and records requests, and signed policy delivered to enrolled desktops. Availability of individual enterprise controls varies by release, so scope a pilot around the controls you need and verify them before rollout. See private AI for business for a pilot plan.

What private AI does not solve on its own

Running a model locally removes one class of risk — disclosure to a model vendor — and leaves others untouched. Be explicit about them when you evaluate any tool, Glyph included.

• Access control still matters. A local index of confidential documents on an unlocked laptop is no more private than the laptop.
• Prompt injection does not care where the model runs. Connectors and browsing extend what an agent can touch; approve them narrowly.
• Model quality is a separate axis. A small local model can be private and still wrong; test it on your own tasks before relying on it.
• Backups and recovery are yours. Local data that exists only on one drive is private and also fragile.
• Compliance is a process, not a checkbox. Fortaify does not claim certifications for Glyph; the Trust Center states what is implemented and what is planned.

Getting started

Install Glyph and verify the installer checksum from the download page. Choose a model sized for your memory — the local LLM guide explains how. Turn on the features you need one at a time, reading what each one sends and where. If you are evaluating for an organisation, start with the readiness checklist and talk to sales about a scoped pilot.