Run Open Models on Claude Desktop: Points: 1 # Comments: 0

What happened

Run Open Models on Claude Desktop

Ollama announced on August 25, 2026 that Claude Desktop can now be configured to use Ollama as a third-party gateway provider, letting users run open models directly through the Claude interface without leaving the application. The change requires no manual configuration file editing; users enable the connection with a toggle inside the Ollama settings screen, and Ollama handles the gateway setup automatically.

This article covers how the integration works, what it means for privacy, and which decisions a developer or privacy-conscious user should verify before relying on the setup. If you use Claude Desktop for daily coding or research and have been curious about swapping in locally hosted models, this is a direct guide to what changed and what you should check.

How it works

Ollama shipped support for Claude Desktop as a third-party gateway consumer, announced on August 25, 2026. The integration lets Claude Desktop route its model requests through Ollama rather than directly to Anthropic's API. Users open Ollama, navigate to settings, find the Claude toggle, switch it on, and Ollama configures the gateway automatically. Turning the toggle off restores whatever Claude setup was in place before, so the change is reversible in a single click.

Once connected, the Claude Desktop interface gains access to both locally running models and models hosted on Ollama's cloud. This means a user who has already pulled open-weight models to their machine can select those models inside Claude Desktop without leaving the application or switching to a separate chat interface. The integration is positioned as additive: Anthropic models remain available and can be swapped back at any time.

Ollama acts as a third-party gateway that sits between Claude Desktop and the model backend. When the toggle is enabled, Ollama intercepts the requests Claude Desktop would normally send to Anthropic and reroutes them to whichever model the user has selected in Ollama's settings. Both local models already downloaded to the user's machine and cloud-hosted models available through Ollama's service can be selected, giving the user a single settings page to manage the choice.

Why it matters

The model selection happens at the per-use level inside Ollama's settings screen. Users navigate to the Settings page to configure a local model, and the Ollama toggle inside Claude becomes the switch that determines whether local or Ollama cloud models are in play or whether the session reverts to native Anthropic. Telemetry is disabled by default, which affects what data the application collects during model inference, though the two paths, local and Ollama cloud, carry different data-handling implications that a builder should verify separately.

Claude Desktop has built a large user base among developers who rely on it for coding and research tasks. Until this change, those users had no supported path to substitute a locally running open model for the Anthropic-hosted one within the same interface. Requiring a separate application to use a local model creates friction, and that friction has historically pushed privacy-sensitive workflows toward third-party wrappers that may not be maintained or audited. A first-party toggle from Ollama removes that friction while keeping the user inside a single, known application.

Who is affected

Ollama's Zero Data Retention policy applies across all of its models and services, both cloud-accessed and locally run. When a model runs purely on the local machine, prompts do not leave the device at all, because the gateway and the model backend are both resident on the same hardware. This combination, Claude Desktop's familiar interface and Ollama's data posture, gives teams a documented baseline to point at when arguing for local inference in a compliance review, which was harder to articulate with informal workarounds.

The most immediate audience is developers who already have Ollama installed and use Claude Desktop as their primary coding assistant. For this group, the change is a drop-in expansion: no new tool is required, and the cost of trying it is limited to the time spent flipping the toggle and selecting a model. Researchers and writers who use Claude Desktop for long-form work and have privacy concerns about sending sensitive material to a hosted API also gain a practical option that did not previously exist within the official application.

Organizations that restrict outbound API calls for compliance reasons are a secondary audience. A team that can approve Ollama's self-hosted path but cannot approve direct Anthropic API traffic now has a documented mechanism to run Claude Desktop inference entirely on internal hardware. Builders embedding Claude Desktop into controlled environments should verify whether Ollama's cloud model option routes data through Ollama's servers, since that path would still involve external traffic, and should test the toggle behavior before rolling it out to end users.

What to watch next

The integration currently routes all Claude Desktop traffic through Ollama, but the actual model selection still happens within Ollama's settings. A builder deploying this in a team environment should confirm that model selection persists across restarts and is not reset by Claude Desktop updates. Anthropic has not announced any changes to Claude Desktop itself in connection with this release, so it is worth watching whether Claude Desktop adds a native model-routing UI that could supersede or conflict with the Ollama gateway approach.

Ollama's announcement is light on specifics about which open models work best as drop-in replacements for the Anthropic-hosted models in coding and reasoning tasks. Builders who want to use this integration for production-adjacent work should run their own evaluations against the task categories they rely on most before switching. Whether Anthropic extends its official third-party gateway program to other local inference providers is also worth monitoring, since the existence of the Ollama path signals that Claude Desktop's architecture already supports the pattern.