LM Studio, the app that brought open-source models to thousands of laptops, is launching Bionic: a full-featured agent for coding, research, and document work. The key difference is you can run it entirely locally, or scale to larger open models in the cloud when needed, without LM Studio retaining your data.

Bionic works on local codebases (inspect, edit, debug with inline diffs), documents and spreadsheets (generate, organize, summarize), and includes a voice keyboard with local transcription that works in any app. The choice between local and cloud stays yours: local models run on LM Studio’s runtime, cloud models (like GLM 5.2 and Kimi K2.7 Code) run on LM Studio Secure Cloud with a Zero Data Retention promise.

Why this matters. If you wanted an agent without sending data outside, options were limited and immature. Bionic closes that gap with an integrated interface, quality local transcription (Mistral AI’s Voxtral), and the option to scale to the cloud only when the task demands it. For anyone working with sensitive data or simply wanting to control costs, it’s the first serious alternative to cloud-only agents.

Getting started. Download LM Studio Bionic (a separate app from classic LM Studio). For cloud models, you need an LM Studio account with active billing. Everything else — local, voice transcription, file editing — is free.

The details.

Context.

LM Studio started as a local runtime for open-source models: download a model from Hugging Face, run it on your hardware, and get an OpenAI-compatible API without sending anything out. It built a solid community among people wanting privacy, those experimenting with open models, and anyone who can’t or won’t pay per-use. The limitation was it stayed an infrastructure piece: you had to build on top of it.

Bionic is the next step: a complete agent app that uses that same local infrastructure, but also gives you the option to go to the cloud when the task is too heavy for your laptop. The core promise is Zero Data Retention: even when you use LM Studio’s cloud models, requests are processed and discarded, not retained for training or anything else.

What it does, in detail.

Bionic works with projects: a Code project points to a local folder, and Bionic can read code, explain it, modify it (with inline diffs you review before applying), and do agentic search through files to understand how something works. Recommended models for coding are GLM 5.2 and Kimi K2.7 Code, both open-source and available in the cloud.

A Work project loads documents, PDFs, slides, spreadsheets: Bionic processes them in a sandboxed environment, can generate new files from scratch, organize local directories, and bring external context with native web search. The interface shows in-app previews for supported file types (the list is growing), and every change goes through automatic checkpoints, so you can review or roll back.

The voice keyboard is a cross-project feature: press a shortcut from any app, speak, and Bionic transcribes where your cursor was. Transcription is completely local and uses Voxtral, Mistral AI’s multilingual model. No external APIs needed, nothing leaves your computer.

The implications.

For anyone building agents on LM Studio with custom scripts, Bionic removes that work: the interface is already there, the sandbox is there, checkpoint management too. For anyone using Claude Code or other cloud agents but worried about privacy, Bionic offers an alternative with the same UX, where data stays local by default.

The business model is clear: no recurring cost if you use only local, pay-per-use billing (LM Studio Secure Cloud) if you switch to large models in the cloud. The choice stays with the user, for each task.

Limitations and open questions.

Bionic just launched. We don’t yet know how well it scales on large projects (codebases with thousands of files, complex repositories), or how stable the sandbox is on non-standard files. The list of formats with in-app preview is growing, so some document types might still need an external app.

The Zero Data Retention promise applies to LM Studio Secure Cloud, but we have no external audits confirming it: it’s a statement, not a certification. For anyone needing to prove compliance, more is needed.

Finally, agent quality depends on the underlying models. GLM 5.2 and Kimi K2.7 Code are competitive on coding, but aren’t yet at the level of GPT-5.6 Sol or Claude Fable 5 on more complex tasks. Bionic gives you the choice, but choice matters: if the model isn’t up to the task, the interface won’t fix it.