update
May 3, 2026
By Teun
Ollama v0.23.0 adds Claude Desktop support
Ollama v0.23.0 pre-release adds support for launching Claude Desktop through Ollama Launch. The update also surfaces featured models in the app, fixes a Windows gateway timeout issue, and improves Metal startup handling on macOS.
Ollama has released v0.23.0 as a pre-release, and the main change is support for Claude Desktop through Ollama Launch. According to the release notes, users can now launch Claude Desktop with ollama launch claude-desktop, while Claude Code remains available in the Claude Desktop app. Claude Code on the terminal can still be reached from the CLI with ollama launch claude.
The update expands how Ollama users can interact with Anthropic’s Claude tools from a local setup. Ollama said Claude Cowork and Claude Code are both supported within the Claude Desktop app, which means the desktop interface can now act as a central place to start those workflows through Ollama Launch.
⚡ New to this?
This news matters because Ollama is adding a simpler way to start Claude tools from its desktop and command-line flow. Claude Desktop is the app users open on their computer, while the CLI is the command line interface, the text-based terminal many developers use. The Windows and macOS fixes also matter because local AI tools often depend on networking and GPU startup behaving correctly.
🦞 OpenClaw angle
If you are packaging local AI workflows for teammates, test ollama launch claude-desktop and ollama launch claude as part of your setup docs so users have one documented path for desktop and terminal use. If you support Windows clients, re-check any local gateway or loopback assumptions, since this release specifically changes the app to enforce IPv4 loopback. On macOS systems with Metal-backed acceleration, add a startup test that catches kernel compilation failures early so users get a clear error instead of a broken launch.
The release also changes how the Ollama app presents model recommendations. According to the changelog, the app now surfaces featured models from server-driven recommendations, which suggests the model list can be influenced by data coming from Ollama’s servers rather than only by a static local list.
There are also two fixes aimed at stability. Ollama said it fixed an OpenClaw gateway timeout on Windows by enforcing IPv4 loopback. That kind of issue usually points to a local networking problem between components on the same machine, and the fix is meant to make Windows behavior more reliable.
On macOS, Ollama said it hardened Metal initialization so it can handle ggml kernel compilation failures more gracefully. Metal is Apple’s graphics and compute framework, while ggml is a machine learning library used by a number of local AI tools. In practical terms, the change is meant to reduce crashes or hard failures during startup when GPU-related compilation does not go as expected.
The release notes also credit a new contributor. @UniquePratham made their first contribution in #15726, according to the changelog.
Ollama described v0.23.0 as a pre-release, and the full changelog compares v0.22.1 to v0.23.0.
