Set up an offline AI IDE in 2026: what actually works

Most tools marketed as "local AI" still phone home. They call telemetry endpoints, check for license validity, or pull model weights from a CDN on first run. If you need a genuine offline AI IDE that works with no internet, the list of options is shorter than most articles suggest.

This guide covers what "offline" actually requires, what hardware you need, and the two realistic setups that work as of 2026.

What "offline AI IDE" actually requires

A tool is genuinely offline only if all three of these are true at runtime:

1. The model runs locally. No API calls to OpenAI, Anthropic, or any hosted endpoint. The model weights live on your machine.
2. The IDE works without an internet connection. No license check, no telemetry, no CDN asset fetch. Cold-start works on a plane.
3. No background egress. No update pings, no crash reporters phoning home, no "anonymous usage statistics."

Most tools meet criterion 1 (you can point them at a local model) but fail on 2 or 3. VS Code itself checks for updates by default. Many extensions send telemetry. "Local mode" in some tools just means the model is local; the app still requires an active session.

Hardware requirements for offline AI coding

The model runs on your machine, so your hardware defines the ceiling. A full GPU guide for local AI is here, but the short version:

CPU inference works but is too slow for interactive coding. 4 GB VRAM is the absolute floor; expect 3-8 tokens per second with small models. Use the VRAM calculator to check whether your GPU fits a specific model and quantization level.

Three categories of tool

Tools that can run offline fall into three categories:

• Standalone desktop apps that ship everything in one binary. These have the best offline story because there is no VS Code host process phoning home separately.
• VS Code forks like Cursor. The fork ships with AI features, but the underlying VS Code base has its own telemetry and update behavior to manage. Cursor is in this category but leans cloud-first. Local mode exists but requires an account.
• Extensions like Continue.dev or Cline. These run inside VS Code or a compatible editor. They delegate the model call to a local runner like Ollama, but the host editor is responsible for the offline behavior.

For strict offline use, standalone apps and properly configured extensions are the two realistic paths. VS Code forks involve more configuration to fully lock down egress.

Option 1: Bodega One (all-in-one, fully offline)

Bodega One is a standalone Electron app: Monaco editor, AI chat, and an autonomous coding agent in one installer. It is built to run air-gapped by default. There is no bundled cloud model, no account requirement to start, and an optional air-gap mode that blocks all network egress at the OS level with nine enforcement layers.

The Bring Your Own LLM (BYOLLM) setup covers 10+ provider presets. For offline use, Ollama is the recommended runner.

Setup: Bodega One with Ollama

1. Install Ollama from ollama.com. Pull the model you want:
   ollama pull qwen2.5-coder:14b-instruct-q4_K_M
2. Install Bodega One from bodegaone.ai/pricing. Windows, macOS, Linux installers available.
3. Open Settings → Providers. Select Ollama from the preset list. The base URL defaults to http://localhost:11434.
4. Select your model from the dropdown (auto-populated from your Ollama library).
5. Enable air-gap mode (optional) under Settings → Privacy. This activates OS-level firewall rules and blocks all outbound connections from the app process.
6. Disconnect from the internet and test. Open a project and run an agent task. If the firewall rules are active, the audit log in Settings → Air-gap shows zero outbound attempts.

The agent runs a 5-step verification pass on every task: it checks its own output against the original spec, runs compile gates, and does a full structural sweep before marking anything complete. See the full Ollama setup guide for model-specific configuration notes.

Option 2: Continue.dev + Ollama (free, VS Code-based)

Continue.dev is an open-source extension for VS Code and JetBrains. It connects to any OpenAI-compatible endpoint, including Ollama running locally. The extension itself does not require an account and does not send your code to any server. VS Code's own telemetry is separate and can be disabled in settings.

Setup: Continue.dev with Ollama

1. Install Ollama and pull a model: ollama pull qwen2.5-coder:14b-instruct-q4_K_M
2. Install the Continue extension from the VS Code marketplace.
3. Open the Continue config file (~/.continue/config.json) and add:
   { "provider": "ollama", "model": "qwen2.5-coder:14b-instruct-q4_K_M" }
4. Disable VS Code telemetry: telemetry.telemetryLevel: "off" in VS Code settings.
5. Test offline: disconnect, restart VS Code, open a file, and trigger a completion.

Continue.dev supports chat, autocomplete, and inline edits. It does not have an autonomous agent loop or a built-in verification layer. For a full comparison of what each tool does and does not include, see AI coding tools that work completely offline (2026).

How to verify nothing is leaving your machine

Claiming a tool is offline is not the same as confirming it. To verify:

• macOS / Linux: Run sudo lsof -i -n -P | grep ESTABLISHED after disconnecting from the internet and launching the IDE. Any external connections will show up.
• Windows: Use netstat -ano from an elevated prompt and cross-reference PIDs with Task Manager.
• Bodega One: The air-gap audit log records every outbound attempt the app makes (or does not make). It is readable in plain text under Settings → Air-gap → Audit Log.

If you see connections to Ollama's localhost address (127.0.0.1:11434), that is expected. Those are local IPC calls to the model runner, not external egress.

Which setup should you use?

Use Continue.dev if you want a free setup, are already in VS Code, and do not need an agentic loop. It works, it is actively maintained, and configuration takes under 10 minutes.

Use Bodega One if you want a standalone app with a full IDE, agent mode, and 9-layer air-gap enforcement. The one-time price is $79 Personal (2 machines) or $149 Pro (5 machines). No subscription, no cloud dependency. Join the waitlist for full launch. The beta is live now.