Run Claude Desktop Offline: Ollama v0.23.0 Local AI

Claude Desktop can now run entirely offline on your own computer — no cloud connection required, no data sent to remote servers. Ollama v0.23.0's native Claude Desktop support makes this possible for developers handling sensitive code, businesses with strict data policies, and anyone concerned about AI privacy, removing the biggest barrier to running Claude as a local AI model without relying on Anthropic's servers.

Claude Desktop Goes Offline: Local AI Without the Cloud

Until this release, using Claude meant every prompt you typed traveled to Anthropic's servers, got processed there, and returned to you over the internet. Ollama v0.23.0 changes the architecture — the underlying plumbing of how the software connects — by enabling Claude Desktop (Anthropic's official desktop client) to run against a locally hosted model instead of the cloud API (the remote connection that normally handles every request).

Two operational modes are now supported:

• Claude Desktop App — Full graphical interface with both Claude Cowork (a structured collaborative workspace for AI-assisted sessions) and Claude Code (an AI-powered coding assistant) running entirely on your own machine
• Terminal / CLI mode — Direct command-line access to Claude Code for developers who prefer working in a text shell rather than a GUI

Ollama acts as the inference backend — think of it as the engine under the hood that actually runs the AI model on your local CPU or GPU (graphics processing unit, the chip originally designed for video games that also excels at AI calculations).

Two Commands to Launch Claude Desktop Offline

The entire setup comes down to two commands depending on how you prefer to work:

# Launch the full Claude Desktop graphical app
ollama launch claude-desktop

# Launch Claude Code in your terminal / command line
ollama launch claude

If you already have Ollama installed, upgrading to v0.23.0 is the only step required. If you are new to Ollama, it is a free, open-source tool that manages local AI models — think of it as an app store for AI that runs entirely on your own hardware, with no subscriptions or cloud accounts needed. The setup guide on this site walks through the full installation from scratch.

Five Releases in Nine Days: A Signal of Intent

The Claude Desktop integration did not arrive in isolation. Between April 24 and May 3, 2026, the Ollama team shipped 5 separate releases — an unusually aggressive cadence that points to a coordinated product push rather than routine maintenance. Many mature open-source projects release updates every few weeks or months. Ollama did it nearly every other day.

The version timeline

• v0.21.2 (April 24) — Base stability work
• v0.22.0 (April 28) — Added NVIDIA Nemotron 3 Omni and Poolside Laguna XS.2 models
• v0.22.1 (April 30) — Renderer optimizations for Gemma 4 thinking and tool-calling
• v0.21.3-rc0 — Release candidate (a pre-final test version distributed before the official launch)
• v0.23.0 (May 3, 2026) — Native Claude Desktop integration plus Windows and macOS stability fixes

The pattern suggests deliberate coordination: Ollama was expanding its model library, hardening platform stability on both Windows and macOS, and landing the Claude Desktop launch all within the same 9-day window. This positions Ollama as the primary local inference runtime (the software layer that runs AI models on personal hardware) for Claude — distinguishing it from alternatives like LM Studio or vLLM, which lack direct Claude Desktop integration.

Two New Local AI Models Added to Ollama's Library

Alongside the Claude integration, v0.22.0 added two models worth knowing about:

• NVIDIA Nemotron 3 Omni — A multimodal reasoning model (one that can process both text and images, not just text alone) from NVIDIA, optimized specifically for on-device inference where cloud GPU resources are unavailable
• Poolside Laguna XS.2 — Billed as the first open-weight coding model from Poolside, a Paris-based AI company specializing in software development AI. "Open-weight" means the model's parameters — the millions of numerical values that define how the AI thinks — are publicly available, so anyone can run, inspect, or fine-tune them without paying a license fee

Ollama now surfaces these as "featured models" based on server-driven recommendations that update independently of the app itself — meaning new model suggestions can appear without requiring you to upgrade Ollama, a useful separation of model discovery from software versioning.

Platform Fixes That Cleared the Path for Offline Launch

Two specific technical issues were resolved in v0.23.0, both of which had been causing real-world failures on users' machines:

• Windows gateway timeout — The OpenClaw gateway (Ollama's integrated web search plugin) was timing out on Windows due to a conflict between IPv6 and IPv4 routing (two different versions of the internet address system that Windows handles differently). The fix enforces IPv4 loopback — the local address a computer uses to talk to itself, separate from any internet connection — resolving the intermittent timeouts
• macOS Metal initialization crash — On Apple Silicon Macs (M1 through M4 chips), ggml kernel compilation — the low-level process that optimizes AI model performance for Apple's Metal GPU framework (the software layer Apple uses to accelerate AI and graphics calculations) — could fail silently and crash Ollama with no useful error message. The fix adds graceful error handling so the app recovers and surfaces a clear diagnostic instead of crashing

These targeted fixes suggest the Ollama team was actively triaging stability reports ahead of the Claude Desktop launch, treating the integration as a high-stakes release rather than a routine patch.

Who Should Run Claude Offline — and What Hardware You Need

Running Claude locally via Ollama makes the most practical sense for these groups:

• Software developers working with proprietary or client-confidential codebases that cannot be sent to external cloud services under their contracts
• Healthcare, legal, and finance teams subject to data residency requirements — regulations that specify where data must physically be stored and processed — which restrict the use of cloud AI on sensitive documents
• Power users on paid Claude plans who want to route routine tasks (summarizing, drafting, Q&A) through a free local model to reduce their monthly costs
• Students and researchers in areas with unreliable or expensive internet who need AI tools that work completely offline

The practical constraint is hardware. A Mac with Apple Silicon (M1 through M4) or a Windows or Linux PC with at least 16GB of RAM handles the smaller Claude model variants comfortably. Larger, more capable variants typically require 32GB of RAM or more. For everyday tasks — writing help, code review, summarizing documents — the smaller local variants perform surprisingly well against their cloud counterparts.

If you have never installed Ollama or run a local AI model before, the beginner's guide on this site walks through the full process from scratch with no prior technical experience needed. Once Ollama is running on your machine, launching Claude Desktop is a single command away — and entirely free to try.

Related Content — Get Started | Guides | More News