Claude Cowork + Local Models: Free Setup in 5 Minutes

Want to run Claude Cowork with local models — completely free, offline, and private? Since January 2026, Ollama v0.14 ships native Anthropic Messages API compatibility, which means Claude's agentic desktop tool can now talk directly to open-source models running on your hardware. No API key. No subscription. No data leaving your machine.

This guide covers everything: installation, configuration, model selection, performance benchmarks, limitations, and a full comparison chart between cloud Claude and local inference. Whether you're a developer concerned about privacy or someone who wants unlimited Claude Cowork usage at zero cost — this is the definitive setup guide for 2026.

#What Is Claude Cowork?

Claude Cowork is Anthropic's agentic desktop tool that brings Claude Code's capabilities to Claude Desktop for knowledge work beyond coding. Instead of responding to prompts one at a time, Claude can take on complex, multi-step tasks and execute them on your behalf — formatting documents, organizing files, synthesizing research, and automating workflows.

Key Capabilities

• Multi-step task execution: Describe an outcome, step away, and come back to finished work
• File system access: Read, write, and organize files on your computer
• Scheduled tasks: Automate recurring work (cloud-only feature)
• Projects: Persistent workspaces with their own files, links, instructions, and memory
• Plugins: Extend functionality with skills, connectors, and sub-agents
• Computer Use: Control desktop apps by seeing, clicking, and typing

Cowork runs directly on your computer in an isolated VM, giving Claude access to files you choose to share. Code executes safely in sandboxed environments while Claude makes real changes to your files.

#Why Use Local Models with Claude Cowork?

Running Claude Cowork against cloud APIs costs money and sends your data to external servers. Here's why local models change the equation:

The tradeoff is clear: local models trade speed for privacy, cost savings, and unlimited usage. For many workflows — especially those involving sensitive code, proprietary documents, or air-gapped environments — that tradeoff makes perfect sense.

#Prerequisites & Hardware Requirements

Before setting up local models with Claude Cowork, ensure your system meets these requirements:

Software Requirements

• Ollama v0.14.0+ (required for Anthropic Messages API compatibility)
• Claude Code CLI installed via curl -fsSL https://claude.ai/install.sh | bash
• macOS 13+, Windows 10+, or Linux (Ubuntu 20.04+ recommended)

Hardware Requirements

#Step-by-Step Setup: Ollama + Claude Code

Step 1: Install Ollama

macOS (Homebrew):

Linux:

Windows: Download from ollama.com

Verify installation:

Step 2: Pull a Local Model

Choose a model with tool calling support (required for Claude Code's agentic features):

Step 3: Install Claude Code

macOS/Linux:

Windows:

Step 4: Connect Claude Code to Ollama

Fastest method — one command:

This automatically sets ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL, and launches Claude Code pointed at your local Ollama instance. Select your model from the list and hit Enter.

Manual method — explicit environment variables:

Or inline without modifying your shell profile:

Step 5: Verify Connection

Once Claude Code launches, try a simple command:

If the model reads files and responds with actual file listings (not just describing what it would do), tool calling is working correctly.

#Setup with LM Studio

LM Studio provides a graphical interface for managing local models:

1. Download LM Studio from lmstudio.ai
2. Search and download GLM-4.7-Flash or Qwen3-Coder
3. Go to Local Server tab → Start Server (default port: 1234)
4. Configure Claude Code:

#Best Local Models for Claude Cowork

Top recommendation: GLM-4.7-Flash. Its mixture-of-experts architecture means only 3B parameters activate per token despite being a 30B model. That translates to fast inference on modest hardware (16GB RAM) while maintaining 128K context and strong tool-calling abilities (79.5% on agent benchmarks).

#Free Cloud Models via Ollama

Don't want to run inference locally? Ollama also proxies free cloud models with generous rate limits:

These models run on remote infrastructure but use the same Ollama interface. Your code still goes to external servers (not truly private), but it's free and significantly faster than local inference.

#Full Comparison: Cloud Claude vs Local Models

#Performance Benchmarks

Real-world numbers from published benchmarks comparing local and cloud inference:

Token Throughput

Real-World Task Timing

Coding Quality Scores (50-task benchmark)

#Cost Analysis

Breakeven point: A heavy Claude Max user ($200/month) recoups GPU investment in just 2.5 months. Even Claude Pro users ($20/month) break even within 6-8 months if they already own capable hardware.

#Limitations of Local Models

Be realistic about what local models cannot do:

• Slower inference (3-68x): Simple tasks take 5x longer. Complex repo analysis can take 68x longer than cloud Claude.
• Lower edit accuracy (70-80% vs 98%): Local models produce patches with wrong line numbers, bad whitespace, and mismatched context. Over a 50-edit session, you'll spend more time fixing broken patches than writing code.
• Weaker multi-file reasoning: Cloud Claude excels at understanding relationships across large codebases. Local models degrade significantly with complexity.
• Tool calling reliability: Not all models support tool calling. Without it, Claude Code becomes a plain text generator that describes actions instead of executing them.
• No scheduled tasks: Recurring automated work only runs with cloud Cowork.
• No Computer Use: Desktop control (clicking, typing in apps) requires cloud Claude.
• No plugins: Most Cowork plugins require cloud infrastructure.
• Context window limits: Local models typically max at 128K tokens vs 200K+ for cloud Claude.
• Streaming tool calls require Ollama 0.14.3-rc1+: The stable release may not handle all tool-use scenarios correctly.

#What's Possible with Local Models

Despite the limitations, local models unlock significant capabilities:

• 100% offline development: Write code on planes, in cafes without WiFi, or in restricted networks.
• Complete data privacy: Proprietary code, PII, medical records, defense contracts — nothing leaves your machine.
• GDPR/compliance: Eliminate cross-border data transfer concerns entirely. No DPAs needed.
• Air-gapped environments: Defense, healthcare, and government organizations can use AI coding assistance without network access.
• Unlimited usage: No rate limits, no monthly caps, no throttling during heavy use.
• Custom fine-tuned models: Train models on your codebase for domain-specific assistance.
• Hybrid workflows: Use local for sensitive work, cloud for complex tasks. Switch instantly.
• Zero-cost experimentation: Try different models, approaches, and prompts without watching a billing meter.

#Troubleshooting

"Connection refused" error

• Ensure Ollama is running: ollama serve
• Check port isn't blocked: curl http://localhost:11434/api/tags
• Verify Ollama version: ollama --version (must be 0.14.0+)

Model just talks instead of acting

If Claude Code responds with "I would read the file..." instead of actually reading it, tool calling isn't working:

• Switch to a model with confirmed tool support: GLM-4.7-Flash or any cloud model
• Update Ollama to 0.14.3-rc1+ for streaming tool calls
• Ensure ANTHROPIC_AUTH_TOKEN is set to ollama, not an actual API key

Slow generation (under 5 tok/s)

• Drop to smaller quantization: Q4_K_M instead of Q6_K
• Reduce context: ollama run glm-4.7-flash --num-ctx 32768
• Switch to GLM-4.7-Flash if using a dense model (MoE = faster)
• Consider Ollama cloud models: ollama launch claude --model qwen3.5:cloud

"Role model" request failures

Claude Code tries to use "haiku" for background tasks. Fix by setting the small model override in your Claude Code settings to use the same local model.

#Frequently Asked Questions

Can I use Claude Cowork completely offline with local models?

Yes. Once you've pulled your model via Ollama, everything runs locally. No internet required for inference. However, some Cowork features (scheduled tasks, plugins, Computer Use) are cloud-only and won't work offline.

Is it really free?

Running local models through Ollama is completely free. No API keys, no billing, no subscription. Ollama's cloud models (like qwen3.5:cloud) are also free with generous rate limits. Your only cost for truly local inference is hardware and electricity.

What's the best model for Claude Code with Ollama?

GLM-4.7-Flash is the top recommendation: 128K context, native tool calling (79.5% benchmark), and it runs on 16GB RAM thanks to mixture-of-experts architecture. For Ollama cloud models, Qwen 3.5 and GLM-5 offer frontier-level quality at zero cost.

How much slower is local compared to cloud?

Expect 3-5x slower for simple tasks and up to 68x slower for complex multi-file analysis. The speed gap is the primary tradeoff. However, for many single-file tasks (code explanation, simple edits, documentation), the delay is tolerable (10-20 seconds vs 3-5 seconds).

Can I switch between local and cloud models?

Yes. Use local models for sensitive/private work and cloud Claude for complex tasks. You can switch by simply changing environment variables or using separate terminal profiles.

Does the quality match cloud Claude?

No. Local models score 85-90% of cloud Claude on single-file tasks but significantly worse on multi-file reasoning (50-60% of cloud quality). Edit accuracy drops from 98% to 70-80%, meaning more manual corrections needed.

#The Bottom Line

Claude Cowork with local models is not a replacement for cloud Claude — it's a complement. The ideal workflow in 2026 looks like this:

• Local models → sensitive codebases, unlimited experimentation, offline work, privacy-first environments
• Ollama cloud models → free, faster than local, good quality, acceptable for non-sensitive work
• Cloud Claude → complex multi-file reasoning, scheduled automation, Computer Use, maximum quality

The setup takes 5 minutes. The cost is zero. If you have a Mac with 16GB+ RAM or a GPU with 12GB+ VRAM, there's no reason not to try it. Start with ollama pull glm-4.7-flash and ollama launch claude — you'll be coding with a local AI agent in under a minute.

For more AI coding tools, explore our Claude Opus 4.6 review and our free AI Image Generator.