Three different things get called offline
Before comparing anything, it is worth separating the claims, because a tool can be honest about all three and still not do what you need.
- Offline after setup, if you built the setup. An extension pointed at a model server you installed. Genuinely offline once running, and a project to get there.
- Offline for some features. Chat works against a local model, while completions and codebase search still call home. Common, and the gap usually appears exactly when you are on a plane.
- Offline out of the box, all features. The engine and a model ship with the editor, and every AI surface runs against them. That is the category Onyx Code is in.
What works in Onyx Code with the network off
All of the following run against the bundled llama.cpp engine on localhost, with no account and no outbound request.
- Inline autocomplete, multi-line, with cross-file context and a dedicated fill-in-the-middle model rather than the chat model doing double duty.
- Next-edit prediction, the model that guesses your following edit after you make one.
- Chat about your codebase, including semantic search over a locally built index with local embeddings.
- Agent edits, staged as reviewable diffs you apply or discard.
- Commit messages and pull request descriptions drafted from your staged changes.
- Every privacy control: Air-Gap Mode, the AI Activity Log, the audit PDF export, the prompt sanitizer and
.onyxnoai. None of them are paid features.
The engine uses Metal on Apple Silicon and Vulkan on NVIDIA, AMD and Intel GPUs, with an in-app Auto, GPU or CPU switch, so “offline” does not have to mean “slow”.
What still needs a network, stated up front
Four things, none of which happen while you are writing code.
- The first model download. The app ships with the engine, and the first-run wizard fetches a model sized to your hardware. Do that once, on any connection, and you are set. On a permanently disconnected machine, model files can be copied across.
- Extensions. Installing from the Open VSX registry needs a connection, like any marketplace. Already-installed extensions do not.
- Updates. Auto-update checks are network calls, and Air-Gap Mode blocks them deliberately.
- Pro licence activation. Normal activation is online and takes a moment. For machines that are never online there is an offline activation file you generate elsewhere and copy across. The Free tier needs no activation at all.
Cloud models, if you choose to use them with your own API key, obviously need a connection too. That is opt-in, off by default, and Air-Gap Mode blocks it outright.
How the common options behave with the network off
| Tool | Chat offline | Autocomplete offline | Setup required |
|---|---|---|---|
| Onyx Code | Bundled engine | Dedicated local model | Install the app |
| VS Code with Copilot, local model | Yes, no account needed | Needs an account and connectivity | Install and run a model runtime |
| Continue.dev with Ollama | Yes | Yes, once configured | Extension, runtime, models, config file |
| Cursor | Inference runs on Cursor servers | No | Sign in |
The detail behind each row is on the comparison pages: versus Cursor, versus GitHub Copilot and versus Continue.dev.
How to prove it on your own machine
Do not take the claim on trust. It takes about five minutes.
- Install Onyx Code and let the first-run wizard fetch a model for your hardware.
- Turn on Air-Gap Mode from the Command Palette, or click the status-bar pill. It goes green and reads AIR-GAPPED.
- Switch off Wi-Fi, or pull the cable, so nothing depends on the app being honest.
- Type in a file and wait for a completion. Ask chat a question about a file you have not opened. Ask for an edit across two files and apply the diff.
- Open the AI Activity Log to see every prompt, file read and edit that just happened, and export it as a PDF if you want a record.
For the paranoid version, including the socket check that proves no outbound connection was opened, the documentation has the exact commands.
Questions
Is there an AI code editor that works completely offline?
Yes. Onyx Code bundles the inference engine and downloads a model at first run, after which chat, autocomplete, next-edit prediction, codebase search and agent edits all run locally with the network off.
Do I need a powerful machine?
A recent laptop is enough for autocomplete and codebase questions. The first-run wizard detects your RAM and GPU and recommends a model that fits rather than leaving you to guess, and a small starter model is installed in the foreground so chat works within a couple of minutes while a larger one downloads behind it.
Does it cost anything?
The Free tier has no time limit and includes local chat, autocomplete, next-edit prediction, codebase indexing and every privacy control. Pro is $39 once for three machines and adds unlimited models, project memory and the Plan and Auto agent modes.
Is my code used for training?
No. In local mode nothing leaves the machine to be trained on, and there is no telemetry by default. If you add your own cloud API key, that provider’s terms apply to those calls, which is why the default is local and the choice is per task.
Try the disconnected test
Install it, let it fetch a model once, then turn the network off and see how much still works.
Free tier, no account, no sign-up, no telemetry. macOS, Windows and Linux.