Cursor made AI-assisted coding feel normal, but it is a closed product that routes your code through servers you do not control. Void is the editor a lot of developers have been waiting for: an open-source project that does the same core jobs — chat, inline edits, autocomplete, and an agent mode — while letting you point it at your own API keys or a model running on your own machine. We spent time driving it to figure out whether "privacy-first" is a real difference or a tagline.
The short version: Void is a credible Cursor alternative for developers who care where their code goes and are willing to trade some polish for control. It is not yet a drop-in replacement if you depend on Cursor's fastest autocomplete or its deep codebase indexing. Here is what that trade looks like in practice.
What Void actually is
Void is a fork of VS Code, which is the single most important fact about it. Your existing extensions, themes, keybindings, and settings carry over because the editor underneath is the same one you already know. There is no relearning the file tree, the command palette, or your muscle memory. You install Void, sign in to nothing, and start working.
The project is open-source and backed by Y Combinator, and the AI layer is built around one idea: bring your own model. Instead of paying a single subscription that bundles the editor and the inference, you supply an API key for the provider you want — Anthropic, OpenAI, Google, OpenRouter, and others — and Void talks to it directly. You pay the provider for tokens, and that is the only model cost.
The feature set maps cleanly onto what you expect from this category:
- Chat in a side panel with your code as context.
- Inline edit (the Cmd/Ctrl-K style flow) where you select code, describe a change, and accept or reject a diff.
- Autocomplete for inline suggestions as you type.
- Agent mode that can read files, make multi-file edits, and run through a task with your approval.
None of this is novel on its own. What is different is that you choose the model behind each feature, and you can mix them — a cheap fast model for autocomplete, a stronger one for agentic edits.
The privacy case, and its real limits
The headline claim is that Void has no middle-man server. When you send a chat message, the request goes from your machine straight to the model provider you configured, or stays entirely on your machine if you run a local model through Ollama or LM Studio. Void the company does not sit in the path, does not retain your prompts, and does not train on your code. For anyone who has had a security review flag "third-party code editor with a retention policy you cannot audit," that architecture is the whole pitch.
But be precise about what this buys you, because the marketing word "private" hides a fork in the road.
Using Void with a hosted model does not make your code private from that model's provider. If you point Void at Claude or GPT, your code still travels to Anthropic or OpenAI under their data terms — Void simply stops being an extra party in the chain. You only get true end-to-end privacy when you run a local model, where nothing leaves your machine at all.
So there are two honest stories here. Story one: you use a hosted frontier model and Void removes one intermediary, which is a real but modest privacy gain. Story two: you run a local model like a Qwen or Llama variant through Ollama, and your code never touches the internet — genuine air-gapped AI assistance, at the cost of the local model being weaker and slower than a frontier API. Void is the only one of the two stories that competitors in this space rarely support well, and it is the reason to pick it.
If you are evaluating Void for a regulated environment or a client that forbids sending source to third parties, the local-model path is the one that actually satisfies the requirement. Test it with the specific model you plan to run before you promise anyone it works, because local model quality varies a lot by hardware.
Where it trails Cursor
Void is a younger project, and you feel it. Autocomplete is functional but does not match the latency and uncanny accuracy of Cursor's purpose-built completion model — Cursor invests heavily in a custom autocomplete that Void, relying on general-purpose models through your API key, cannot fully replicate. Codebase-wide retrieval is present but less polished than Cursor's indexing, so on large repositories the agent sometimes needs more hand-holding about which files matter.
You will also hit the normal rough edges of fast-moving open-source software: occasional UI quirks, features that change between releases, and a smaller pool of documentation and community answers when something breaks. If your work depends on the editor being invisible and never surprising you, that friction is real.
A pragmatic setup: use a strong hosted model (Claude or GPT class) for agent mode and inline edits where reasoning quality matters most, and a small fast model — local or cheap — for autocomplete where latency matters more than brilliance. Void lets you assign models per feature, so you are not forced into one choice for everything.
If you would rather not assemble the pieces yourself — pick models, manage API keys, accept the rough edges — the managed alternative Void is chasing is worth a look. Cursor handles the model, the indexing, and the autocomplete tuning for you in exchange for sending your code through its service and paying a flat subscription.
Who should actually switch to Void? Developers who want to keep their code off third-party servers, teams that need a local-model option for compliance reasons, and anyone who likes the idea of an open-source editor they can inspect and fork. Who should stay on Cursor for now? Developers who value autocomplete speed and zero-setup indexing above everything, and who are comfortable with a closed, hosted product.
Void is not finished, but it does not need to be finished to be useful. It is the clearest answer yet to a question more developers are asking: can I get AI in my editor without handing my source to a vendor I cannot audit? With a local model, the answer is finally a straightforward yes.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)