Ollama Speaks Anthropic Now — Here’s How to Run Claude Code for Free
Ollama Speaks Anthropic Now — Here’s How to Run Claude Code for FreeClaude Code thinks it’s talking to Anthropic. It’s talking to your machine. $200/month → $0.
Three environment variables. One command. Claude Code runs against free, local models — and doesn’t know the difference.
In January 2026, Ollama quietly added Anthropic Messages API compatibility. That one update turned Claude Code — Anthropic’s $200/month terminal coding agent — into a free, offline, fully private coding tool. Same agentic workflow. Same file editing, code navigation, multi-turn conversations. Different brain behind it. Your brain. Running on your GPU. Sending nothing to the cloud.
🧠 What's Actually Happening Here — The 60-Second Version
Think of Claude Code as a really smart project manager. It reads your files, plans changes, writes code, runs tests, and edits across your codebase. Normally, every time it “thinks,” it calls Anthropic’s servers — and you pay per token.
Ollama is a local AI model runner. It downloads open-source AI models to your machine and serves them through an API.
The trick: Ollama now speaks the same language as Anthropic’s API. So you tell Claude Code “hey, Anthropic’s server is at localhost:11434” — and it believes you. It sends all its requests to your local Ollama instance instead. Ollama responds using whatever model you picked. Claude Code processes the response like nothing changed.
What you keep:
- Agentic coding workflow (file reading, editing, terminal commands)
- Multi-turn conversation context
- Tool calling and function execution
- Project-wide code navigation
What you trade:
- Peak intelligence (local models aren’t Opus-level — yet)
- Speed on complex multi-file refactors
- Long context reliability above ~64K tokens
For everyday coding — scaffolding, tests, boilerplate, quick edits, debugging — local models handle it fine.
🔧 Full Setup — 5 Minutes, Any OS
Step 1: Install Ollama
| OS | Command |
|---|---|
| Linux | curl -fsSL https://ollama.com/install.sh | sh |
| macOS/Windows | Download from ollama.com |
Verify: ollama --version — need v0.14.0+ for Anthropic API compatibility.
Step 2: Pull a coding model
ollama pull qwen3-coder
Other options depending on your hardware — see the model comparison table below.
Step 3: Launch Claude Code with Ollama
The fast way (Ollama v0.15+):
ollama launch claude --model qwen3-coder
The manual way (any Ollama version):
export ANTHROPIC_BASE_URL="http://localhost:11434"
export ANTHROPIC_AUTH_TOKEN="ollama"
export ANTHROPIC_API_KEY=""
claude --model qwen3-coder
That’s it. Claude Code is now running locally.
Step 4 (optional): Make it permanent
Add to ~/.bashrc or ~/.zshrc:
export ANTHROPIC_BASE_URL="http://localhost:11434"
export ANTHROPIC_AUTH_TOKEN="ollama"
export ANTHROPIC_API_KEY=""
For Windows PowerShell:
$env:ANTHROPIC_BASE_URL = "http://localhost:11434"
$env:ANTHROPIC_AUTH_TOKEN = "ollama"
$env:ANTHROPIC_API_KEY = ""
Pro tip: Test offline by disconnecting WiFi and running a prompt. If it responds — you’re fully local. Nothing leaves your machine.
🏆 Best Models for Claude Code — 2026 Tier List
Not all models are equal. Context window matters. Tool calling support matters. Coding ability matters.
| Model | Size | VRAM Needed | Best For | SWE-bench | Context |
|---|---|---|---|---|---|
| GLM-5 | 744B MoE / 40B active | Multi-GPU or cloud | Best local coding quality | 77.8% | 128K+ |
| qwen3-coder | 480B MoE / ~30B active | 32GB+ RAM | All-around coding + 1M context | 70.6% | 1M |
| gpt-oss:20b | 20B | 16GB+ | Fast local performance | ~65% | 128K |
| GLM-4.7 Flash | 30B | 24GB | Best value/speed tradeoff | 73.8% | 200K |
| devstral-2-small | 24B | 20GB+ | Solid on Apple Silicon | ~68% | 128K |
| qwen2.5-coder:7b | 7B | 8GB | Budget hardware, still decent | — | 32K |
Cloud models via Ollama (no local GPU needed):
| Model | Command | Cost |
|---|---|---|
| GLM-4.7 cloud | ollama launch claude --model glm-4.7:cloud |
Free tier available |
| qwen3-coder cloud | ollama launch claude --model qwen3-coder:480b-cloud |
Free tier available |
| minimax-m2.1 cloud | ollama launch claude --model minimax-m2.1:cloud |
Free tier available |
Pro tip: Ollama recommends models with at least 64K context for Claude Code. Smaller context = more confusion on multi-file tasks.
💰 The Math — What You're Actually Saving
| Scenario | Monthly Cost |
|---|---|
| Claude Code via Anthropic API (Opus 4.5) | $100–$200+ depending on usage |
| Claude Code via Anthropic API (Sonnet 4.5) | $17–$100 |
| Claude Code via Ollama (local) | $0 (electricity only) |
| Claude Code via Ollama (cloud models) | $0–$3/month |
| Claude Code via cheap API providers | $3–$10/month |
Third-party API alternatives can save up to 98% compared to Anthropic’s Opus pricing. DeepSeek V3.2 runs at ~$0.28 per million tokens. Ollama local is literally free.
The real savings compound fast. Heavy Claude Code users burn through tokens during iterative debugging sessions. A single complex refactor can eat $5–15 in Opus tokens. Locally? That same session costs zero — as many iterations as you want.
⚠️ Honest Limitations — What Local Can't Do (Yet)
| Limitation | Reality |
|---|---|
| Complex multi-file refactors | Local models stumble on 10+ file changes — Opus handles these gracefully |
| Long context accuracy | Above ~64K tokens, local models start losing track. Opus maintains coherence at 200K+ |
| Speed | Expect 20–55 seconds for simple prompts on Apple Silicon. Desktop GPUs are faster but still slower than cloud API |
| Reasoning depth | For architecture decisions and complex debugging chains, Opus/Sonnet still have a meaningful edge |
| First-run download | Models are 8–24GB. First pull takes time depending on your internet |
The honest take: Use local models for 80% of your coding — scaffolding, tests, boilerplate, quick fixes, refactoring single files. Switch to Anthropic’s API for the remaining 20% — complex architecture, multi-file changes, gnarly debugging. Your monthly bill drops from $200 to maybe $20.
🛠️ Resources — Everything Referenced
| Resource | What It Is |
|---|---|
| Ollama Claude Code docs | Official setup guide |
| Ollama blog announcement | The January 2026 post that started it all |
| Ollama download | Install Ollama itself |
| Claude Code install | Install Claude Code CLI |
| Ollama model library | Browse all available models |
| llama.cpp | Alternative local inference engine (more control, harder setup) |
| LM Studio | GUI-based local model runner (alternative to Ollama) |
Quick Hits
| Want | Do |
|---|---|
 Fastest setup |
→ ollama launch claude --model gpt-oss:20b |
 Best coding quality (local) |
→ ollama pull glm-5 (needs multi-GPU) |
 Best on 16GB RAM |
→ ollama pull gpt-oss:20b |
 Best on Apple Silicon |
→ ollama pull devstral-small-2 or glm-4.7-flash |
 No GPU at all |
→ ollama launch claude --model glm-4.7:cloud |
 Maximum savings |
→ Local for daily work, Anthropic API for complex-only |
Your code stays on your machine. Your wallet stays in your pocket. Anthropic doesn’t need to know.
!