Claude Desktop Now Runs Local Models Through Ollama

One Toggle Between a Cloud App and Your Own Hardware

Ollama announced official Claude Desktop support on August 25, 2026, and the setup story is the interesting part. You open the Ollama app, select Claude, and turn the integration on. Ollama writes Claude Desktop's third-party gateway configuration for you. Turn it off and your previous Claude setup comes back.

• Ollama configures Claude Desktop's third-party gateway automatically — no hand-edited config files, and the toggle is reversible
• Every model in your Ollama library appears in Claude Desktop's model picker, whether it is running locally on your machine or in Ollama Cloud
• The agent harness stays intact — skills, MCP servers, and file tools continue to work against the open model you select
• Zero Data Retention across Ollama's models and services, with telemetry off by default and prompts not sent to Anthropic when a local model is in use

Why the Model Picker Detail Is the Whole Story

Running open-weight models locally has never been the hard part. Ollama solved that a long time ago, and our comparison of Ollama, vLLM and llama.cpp covers how mature that layer has become. The hard part has been that the good application shell and the local model tended to live in different worlds. You could have the polished agent interface with a hosted model, or you could have a local model behind a plainer front end.

Putting Ollama's models into Claude Desktop's own picker collapses that choice. The application you already know keeps its file tools, its MCP server connections, and its skills, and the inference happens on hardware you own. That is a different proposition from running a local model in a separate chat window, because the value in an agent app is mostly in the surrounding machinery rather than the raw token stream.

What Does This Change for Local AI Hardware?

The practical effect is that a machine sized for local inference becomes considerably more useful. We have covered a steady stream of hardware landing in exactly that bracket, from the Mac mini M6 running local LLM prompts 4.8x faster at $899 to open-weight releases like GLM-5.3-Flash fitting into 128GB of system RAM and Qwen3.8-Flash-Next running in 75GB with no GPU. Each of those made local inference cheaper. None of them made the local model easier to use inside a good agent client.

There is also a cost dimension that is easy to underplay. Agent workloads are token-hungry in a way that ordinary chat is not — a single run that reads files, calls tools, and iterates can burn through far more tokens than a conversation. Pointing that workload at a model running on your own machine changes the per-token cost to zero, which alters the calculus on how freely you let an agent explore.

Local, Cloud, or Both

The integration covers Ollama Cloud as well as local execution, and that combination is more useful than either alone. Small routine work runs locally. Something that needs a larger model than your machine can hold gets pushed to Ollama Cloud without leaving the same picker. Anthropic's own models remain one toggle away for the tasks where they are the right call.

Ollama states a strict Zero Data Retention policy across its cloud and local services and says telemetry is disabled by default. For anyone whose interest in local models is primarily about keeping data on their own hardware, the local path keeps prompts off both companies' servers entirely — the same argument that runs through our wider AI coverage of self-hosted inference.

This is a small feature by the standards of model releases. It is also the kind of plumbing that determines whether open-weight models get used or merely admired, and that makes it worth more than its announcement size suggests.

Sources: Ollama Blog — August 25, 2026; The New Stack — August 2026.