Every line of code you send to Claude, GPT, or Gemini goes through their servers. For many developers and companies, that’s a dealbreaker. Here are the best AI coding setups that keep your code completely private.

Update (April 27, 2026): OpenAI released Privacy Filter, an open-weight model that detects API keys and secrets in text before they leave your machine. Runs locally, Apache 2.0 license. Useful as a pre-flight check in any AI coding pipeline.

Why privacy matters for coding

• Proprietary code — Your company’s source code is intellectual property
• Compliance — GDPR, HIPAA, SOC 2, and other regulations may prohibit sending code to third parties
• Security — API keys, database schemas, and infrastructure details in your code
• Competition — You don’t want your AI-assisted innovations training someone else’s model

If you do use cloud APIs, make sure credentials are handled properly — see our guide on how to secure AI API keys to avoid leaking secrets through your AI toolchain.

Tier 1: Completely local (zero cloud)

Nothing leaves your machine. No API calls, no telemetry, no internet required.

Setup: Ollama + Aider + Continue.dev

# Install Ollama
brew install ollama

# Pull models
ollama pull qwen3.5:27b      # Chat/coding (16GB RAM)
ollama pull codestral:22b     # Autocomplete (12GB RAM)

# Terminal coding
pip install aider-chat
aider --model ollama/qwen3.5:27b

# IDE coding — install Continue.dev in VS Code

Best local models for coding: · Best VPNs for Developers

Model	Size	VRAM	Quality	License
Qwen 3.6-27B	16GB	16GB+	Very good	Apache 2.0
Gemma 4 27B	16GB	16GB+	Very good	Gemma
Codestral 22B	12GB	12GB+	Best autocomplete	MNPL
Qwen 3.5 9B	5GB	8GB+	Good	Apache 2.0
Gemma 4 12B	7GB	8GB+	Good	Gemma

See our Ollama guide for detailed setup and our best AI models for Mac for Apple Silicon optimization.

Hardware recommendations

Budget	Hardware	Best setup
$600	Mac Mini M4 16GB	Qwen 3.5 9B
$1,150	Mac Mini M4 32GB	Qwen 3.5 27B + Codestral
$1,800	Mac Mini M4 Pro 48GB	Qwen 2.5 Coder 32B + Codestral
$300	Used RTX 3060 12GB	Codestral 22B
$800	Used RTX 3090 24GB	Qwen 3.5 27B

See our best GPU for AI and cheapest way to run AI locally guides.

Tier 2: Self-hosted server (your infrastructure)

Run frontier-class models on your own servers. Code stays within your network but you need serious hardware.

Self-hosted options

Model	Hardware needed	Quality
GLM-5.1 (754B)	4x A100 80GB	Frontier (#1 SWE-Bench)
DeepSeek V3 (671B)	4x A100 80GB	Frontier
Kimi K2.6 (1T)	4x A100 80GB (Q4)	Frontier
Mistral Large 2 (123B)	1x H100	Near-frontier
Qwen 3.5 72B	2x A100	Very good

Deployment with vLLM

pip install vllm

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3.5-72B-Instruct \
  --tensor-parallel-size 2 \
  --port 8000

Then point your tools at it:

# Aider
aider --model openai/qwen3.5-72b --openai-api-base http://your-server:8000/v1

# Claude Code (via Anthropic-compatible endpoint)
export ANTHROPIC_BASE_URL="http://your-server:8000/v1"
claude

Tier 3: Privacy-respecting cloud APIs

Some providers offer better privacy guarantees than others:

Provider	Data retention	Training on your data	Privacy
Anthropic API	30 days	No (with API)	Good
OpenAI API	30 days	No (with API)	Good
DeepSeek API	Unknown	Unknown	⚠️
Azure OpenAI	0 days (configurable)	No	Best cloud
AWS Bedrock	0 days	No	Best cloud

Note: Subscriptions (ChatGPT Plus, Claude Pro) may have different data policies than API access. Always check the current terms.

The privacy-first toolkit

Need	Tool	Privacy level
IDE assistant	Continue.dev + Ollama	Full local
Terminal coding	Aider + Ollama	Full local
Autocomplete	Codestral via Ollama	Full local
Complex tasks	Self-hosted GLM-5.1 or Qwen 72B	Your server
Code search	Codestral Embed (self-hosted)	Your server

What you give up

Being honest about the tradeoffs:

1. Quality gap — Local 27B models are good but not Claude Opus good. The gap is ~15-20% on complex tasks.
2. Speed — Local inference is slower than cloud APIs, especially on consumer hardware.
3. Cost — Hardware costs money upfront. A Mac Mini M4 32GB ($1,150) pays for itself vs Claude Pro in ~5 years of subscription savings, but the upfront cost is real.
4. Maintenance — You manage updates, model downloads, and hardware issues.

For most developers, the hybrid approach works best: local models for routine work, cloud APIs (with good privacy policies) for the hardest problems.

FAQ

What’s the most private AI coding agent?

Running Qwen 3.5 27B or Devstral 2 locally via Ollama with Continue.dev or Aider gives you complete privacy — your code never leaves your machine. For cloud options, Azure OpenAI and AWS Bedrock offer zero data retention policies.

Do AI coding tools send my code to the cloud?

Most cloud-based tools (Copilot, Cursor, Claude Pro) send code to remote servers for processing. API access typically has better privacy policies than subscription products. Local models via Ollama send nothing anywhere.

Can I use AI coding tools and stay GDPR compliant?

Yes, by using fully local models or cloud providers with zero data retention (Azure OpenAI, AWS Bedrock). Self-hosted solutions on your own infrastructure give you full control over data residency and processing.

Related: Self-Hosted AI vs API · How to Sandbox Local AI Models · Run AI Offline · Cheapest AI Coding Setup 2026 · Best VPNs for Developers · Ccpa Ai Developers