Ollama for Claude Code: Run Free Local Models (2026 Setup)

How Claude Code runs on Ollama

Claude Code is the agent: the program that reads your files, runs commands and asks before it changes anything. The model is the brain behind it, normally Claude on Anthropic's servers.

Since version 0.14.0 in January 2026 (per Ollama's release notes), Ollama answers in the same message format as Anthropic's API, on your own computer at port 11434. Point Claude Code there and it works with an open model (one anyone can download and run): that is all a Claude Code local LLM setup is. Claude Code itself doesn't change, and it stays under Anthropic's terms whatever model sits behind it.

What you need

Memory decides almost everything: a model must fit in the memory your graphics chip can use, with room for the conversation, or it crawls. You need no Claude plan or API key, and Ollama can install Claude Code for you (or see how to install Claude Code). The best Ollama model for Claude Code is the largest one that fits your memory with room for a 64K context. Where to start with a Claude Code local model:

Hints are what Ollama's launcher shows and downloads come from the Ollama library, both checked October 22, 2026. On a PC, compare the hint with your graphics card's memory. The Mac app is about 214 MB.

Step 1: Install Ollama on Mac, Windows or Linux

The Claude Code Ollama setup starts with the installer from ollama.com/download: an app that runs a small server in the background, plus the ollama command.

1. Mac (macOS 14 or newer): download Ollama.dmg, drag Ollama to Applications and open it; it offers to add the ollama command to your terminal.
2. Windows 10 (22H2) or 11: run OllamaSetup.exe, which needs no administrator rights.
3. Linux: run the install script below; it sets Ollama up as a background service.
4. Check it: open a new terminal window and run ollama -v.

The official one-line installers, for Mac or Linux:

$ curl -fsSL https://ollama.com/install.sh | sh
$ 

On Windows, in PowerShell:

PS> irm https://ollama.com/install.ps1 | iex
PS> 

$ ollama -v
ollama version is 0.40.8
$ 

For a Claude Code Ollama Windows setup, run every command in this guide in PowerShell. On Windows, NVIDIA cards need driver 551.61 or newer, and AMD Radeon cards run through ROCm or Vulkan, according to Ollama's Windows docs.

Step 2: Set the context length to 64K

Context length is how much of the conversation the model can see at once, in tokens (small chunks of text). Claude Code's opening request alone, its instructions plus a description of every tool, was about 20,300 tokens in our runs with no plugins, and users report up to 36,000. That's why Ollama asks for 64K.

Ollama picks a default from your graphics memory, and most computers land below that. On a Mac it counts the share the graphics chip may use, and our 48 GB MacBook Pro's log shows the result: vram-based default context total_vram="37.4 GiB" default_num_ctx=32768, so 32K.

You can raise the Claude Code Ollama context length in four ways, all on the Ollama side:

• The Ollama app (Mac and Windows): open Settings and set Context length to 64k, which the app saves as 65,536 tokens.
• A terminal, with the app closed: run OLLAMA_CONTEXT_LENGTH=65536 ollama serve and leave the window open (what we did).
• A Linux service: run sudo systemctl edit ollama, add the two lines below, then restart the service.
• One model, any system: save a copy of the model with a bigger window and use its new name (here qwen3.8-64k) instead of qwen3.8:27b.

[Service]
Environment="OLLAMA_CONTEXT_LENGTH=65536"

$ sudo systemctl daemon-reload && sudo systemctl restart ollama
$ 

For the copy, put these two lines in a plain text file called Modelfile (Notepad works), then create the model from that folder:

FROM qwen3.8:27b
PARAMETER num_ctx 65536

$ ollama create qwen3.8-64k -f Modelfile
$ 

Then load a model and check what Ollama actually gave it:

$ ollama ps
NAME           ID              SIZE     PROCESSOR    CONTEXT    RUNNER    UNTIL
qwen3.8:27b    47951e4a6661    22 GB    100% GPU     65536      mlx       4 minutes from now
$ 

CONTEXT should be 64K or more and PROCESSOR 100% GPU; a split such as 48%/52% CPU/GPU means part of the model spilled into slower memory. On a Mac, RUNNER mlx means Apple's MLX engine.

Step 3: Start Claude Code with ollama launch claude

Open a terminal in the folder you want Claude Code to work on (Terminal on a Mac, PowerShell on Windows) and run:

$ ollama launch claude
$ 

1. Pick a model. Under "Recommended", the list shows Ollama's cloud models first (they need an ollama.com sign-in), then the local ones with the memory they need, marked "not downloaded" until you have them.
2. Let it download. If the model isn't on your computer yet, Ollama downloads it first. The 18 GB qwen3.8:27b took 6 minutes 43 seconds on our connection.
3. Install Claude Code if asked. If it's missing, Ollama asks "Claude Code is not installed. Install now?" and runs Anthropic's official installer.
4. Trust the folder. A first-time Claude Code asks for a theme and whether you trust the folder, with no login screen. Then the header shows the model that answers: qwen3.8:27b · API Usage Billing.

That yellow line matters. Claude Code knows Claude models' context sizes, not qwen3.8:27b's, so it printed Model "qwen3.8:27b" isn't recognized · context assumed to be 200k tokens, as Anthropic's model docs describe, and would wait far too long to compact (summarize) a long conversation. Start it with a size just under Ollama's and the warning goes away:

$ CLAUDE_CODE_MAX_CONTEXT_TOKENS=64000 ollama launch claude --model qwen3.8:27b
$ 

PS> $env:CLAUDE_CODE_MAX_CONTEXT_TOKENS = "64000"
PS> ollama launch claude --model qwen3.8:27b
PS> 

For that one session, the launcher points Claude Code at Ollama with a placeholder token, so no Claude login is needed, and sends every model slot, helper agents included, to your model. It writes no Claude Code settings: plain claude in a new window is back on your Claude plan. --yes (with --model) skips every question, and anything after -- goes to Claude Code, as in ollama launch claude --model qwen3.8:27b -- --resume.

The manual route: environment variables

To start plain claude instead, set the variables yourself. Ollama's Claude Code docs list the first three; the model lines are what the launcher adds, because some Claude Code helpers otherwise ask Ollama for a Claude model and get "model not found". On Mac or Linux:

$ export ANTHROPIC_BASE_URL=http://localhost:11434
$ export ANTHROPIC_AUTH_TOKEN=ollama
$ export ANTHROPIC_API_KEY=""
$ export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3.8:27b
$ export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3.8:27b
$ export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.8:27b
$ export CLAUDE_CODE_SUBAGENT_MODEL=qwen3.8:27b
$ export CLAUDE_CODE_MAX_CONTEXT_TOKENS=64000
$ claude --model qwen3.8:27b
$ 

On Windows, in PowerShell:

PS> $env:ANTHROPIC_BASE_URL = "http://localhost:11434"
PS> $env:ANTHROPIC_AUTH_TOKEN = "ollama"
PS> Remove-Item Env:ANTHROPIC_API_KEY -ErrorAction SilentlyContinue
PS> $env:ANTHROPIC_DEFAULT_OPUS_MODEL = "qwen3.8:27b"
PS> $env:ANTHROPIC_DEFAULT_SONNET_MODEL = "qwen3.8:27b"
PS> $env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "qwen3.8:27b"
PS> $env:CLAUDE_CODE_SUBAGENT_MODEL = "qwen3.8:27b"
PS> $env:CLAUDE_CODE_MAX_CONTEXT_TOKENS = "64000"
PS> claude --model qwen3.8:27b
PS> 

PowerShell can't hold an empty variable, so the Windows version removes ANTHROPIC_API_KEY, with the same effect. The settings last until you close the window. To confirm either route, type /status: the auth token line should name ANTHROPIC_AUTH_TOKEN and the base URL your own computer on port 11434.

Best Ollama models for Claude Code

The Ollama models for Claude Code that the launcher recommends change over time. On October 22, 2026 the local picks were the three in the table above: all can call tools, which Claude Code depends on (read a file, edit it, run a command), all reach a 256K maximum context and all use the Apache 2.0 license. The cloud picks:

Now the honest part. Terminal-Bench 4.0 is a set of real terminal tasks that Artificial Analysis runs the same way for every model:

GLM-5.3-Flash and DeepSeek V4.1 Flash, both on Ollama's cloud, score about six times what Qwen3.8 27B does, and Claude stays far ahead of all of them. Our AI model leaderboard has the wider picture.

What we saw on a 48 GB Mac

To see the gap on a small, real job, I gave both local models a demo yoga studio's class list and the test prompt from our guide to AI coding agents, which you can run on your own files:

Work only inside this folder. Read [FILE WITH YOUR PRICES OR SERVICES] and build a one-page price list called prices.html that looks good on a phone. Before you change anything, tell me your plan in three bullet points and wait for my OK. When you finish, list every file you created or changed and how I can undo it.

qwen3.8:27b got it right: a clean page, all eight prices exact, an accurate plan, no tool errors. Its first reply took 3 minutes 27 seconds because it read Claude Code's 20,300-token opening request at about 117 tokens a second; then Ollama's cache made each step quick.

qwen3.5:4b read the same request in about 28 seconds and finished in under two minutes, but it's a fast demo, not a reliable worker: right prices, one hard-to-read row, and an em dash in its reply that the folder's CLAUDE.md forbids. On its first run, at the 32K default, its plan priced a class at $29 (the file says $28), after this:

Ollama sent an HTTP 500 error after 5 minutes 22 seconds while the model was still busy with a small side request; Claude Code showed API error · Retrying in 0s · attempt 1/10 and retried by itself. We fixed nothing by hand in any run.

What stops working without Claude

The agent is the same program, so most of Claude Code keeps working. What depends on Anthropic's servers or a Claude login changes:

Sources: Anthropic's feature availability page and Ollama's Anthropic compatibility list.

Ollama cloud models for Claude Code

The Claude Code Ollama cloud route uses models ending in :cloud, which run on Ollama's servers: no download, no strong graphics chip, much higher scores. The catch: they need an ollama.com account (ollama signin), and your prompts leave your computer.

$ ollama launch claude --model gemma4:cloud
$ 

The free plan has starter usage credits, in an amount Ollama doesn't publish, for a few starter models such as gemma4:cloud. Pro is $20 a month ($200 a year) with $60 of usage a month, and Max $100 with $300, per Ollama's pricing. Ollama says it doesn't store, log or train on cloud prompts, and hosts mainly in the United States.

ollama.com/pricing

What leaves your computer

Troubleshooting Claude Code with Ollama

Most Claude Code Ollama API errors come down to Ollama being off or busy, a wrong model name, too little context or too little memory.

The login and connection rows follow Anthropic's gateway troubleshooting docs; for the rest, Ollama's log on a Mac is ~/.ollama/logs/server.log.

Other ways to run Claude Code on a local model

Details are from LM Studio's and OpenRouter's docs. For an agent built to work with any model, read Claude Code vs OpenCode.

What to do next

Install Ollama, set the context to 64K and give qwen3.8:27b (or qwen3.5:4b on a smaller computer) our small job on a copy of a real folder. Time it and check the result yourself before trusting it with more. If the gap to Claude is too wide for your work, compare plans in our Claude Code pricing guide, and if you want a guided path through real projects, the Claude Code course is open for early access.