A desktop workspace for the models you run with Ollama — and a runtime of its own.

What Ollama is, and what it is not

Ollama is an open-source runtime and command-line tool that downloads open models into a local store and serves them through an HTTP API bound to the loopback interface on port 11434. It is excellent at what it does: one command to pull a model, one to run it, a stable API for scripts and applications. It is not a workspace. There is no document knowledge base, no personas or memories, no image or voice tools, and no policy or identity layer — those come from whatever you put in front of it.

What Glyph does instead

Glyph is the end-user application: chat and projects, a knowledge base over your files, personas and memories, local image generation, transcription and speech, agent tools and connectors, and GlyphNet for using the desktop from your phone. It bundles the runtime, so installing Glyph is the whole setup — choose a model from the Hugging Face-backed library and it downloads, loads with the right GPU acceleration, and runs. On Apple silicon it also runs MLX models, which Ollama does not.

If you already have GGUF files from Ollama-era experiments or other tools, put them in Glyph’s models folder and they appear in the library; Glyph reads the model name from the GGUF header. Ollama’s own store keeps models as content-addressed blobs, so for a model you pulled with Ollama the simplest path is to download the same repository inside Glyph.

Where Ollama plugs in: the Glyph Private server

For a team, the question changes from “which app on my laptop” to “where does the shared model run and who may use it.” Glyph Private answers that with a customer-operated server whose gateway routes requests to configured backends, and Ollama is one of them. The adapter uses Ollama’s native API — model listing, pull with progress, streamed chat, and embeddings — and the Docker Compose deployment includes an optional Ollama profile so a single host with a GPU can serve the team.

Employees’ desktops enrol with the server, sign in through your identity provider, and see the models the administrator published as aliases. Requests pass through authentication, policy, DLP rules, and a usage record before reaching Ollama. Clients never talk to the Ollama port directly, which is the security property a raw Ollama server on a network cannot give you.

Using Glyph as the endpoint other tools call

Many editors, scripts, and agent frameworks are configured to talk to an OpenAI-compatible endpoint at an Ollama or LM Studio address. Glyph runs its own OpenAI-compatible local API — chat completions, completions, embeddings, models, and a Responses API shim — on the loopback interface (port 1234) by default, with an explicit option to expose it on your network. Point the tool’s base URL at Glyph and it uses whatever model Glyph has loaded, on the same hardware, with the same privacy boundary.

Running both on one machine

Nothing prevents Glyph and Ollama from coexisting. They use different ports and different model stores. The practical constraint is memory: two runtimes each holding a model will contend for the same GPU or unified memory, so unload one before loading a large model in the other. If you keep Ollama for scripting and Glyph for daily work, choose a small model in Ollama and let Glyph own the big one, or use Glyph’s local API for the scripts too.

Private by configuration, in either setup

A local runtime keeps inference on owned hardware, and that is true for both tools. What decides privacy is the rest of the configuration: whether the API is bound to loopback or exposed to the network, whether web search or cloud providers are enabled, and which connectors an agent may call. Glyph makes those choices visible in the interface; on a Glyph Private deployment an administrator makes them once in policy for every enrolled desktop.