LM Studio Bionic: What the AI Agent for Open Models Can Do

On July 16, 2026, the LM Studio team shipped Bionic — a standalone app that turns open models into a working agent. It isn't a new mode inside the familiar chat, it's a second product: the developers themselves call it the next generation of LM Studio for agentic work. Here's what's actually in it, and where "local" stops.

How Bionic differs from regular LM Studio

Classic LM Studio is a shell for running models: a chat, Hugging Face model search, loading settings, a local server, and the role of an MCP host. You talk to the model, or hand it to another app over API.

Bionic solves a different problem — carrying work through to a result. It reads a codebase, proposes edits, goes through documents, and searches the web. Format-wise, it's closer to a desktop AI agent than a chat: the agent decides on its own which files to open and which tools to call.

Bionic installs separately, for macOS, Windows, and Linux.

Two kinds of projects

Work is organized around projects, and there are exactly two kinds.

Code Projects — work inside a local codebase. The agent gets tools for files, search, Git, and shell, can search across the project, and shows edits as inline diffs rather than a paragraph saying "replace this function." This is the same class of task as an AI coding agent.

Work Projects — research, writing, analysis, and documents: PDFs, presentations, spreadsheets. Bionic works with these in an isolated environment, and can assemble a new document or go through and summarize existing ones.

Multiple sessions can run inside a project, so different tasks don't interfere with each other and can run in parallel.

There's also local speech transcription: LM Studio uses the multilingual Voxtral model, and you can dictate into any app, not just Bionic.

Where the model computes

Bionic doesn't have one answer: the compute location is chosen per session, and there are three options.

1. A local model on your computer — through the llama.cpp and MLX engines, as in regular LM Studio.
2. An open frontier model in LM Studio Secure Cloud — the company's own cloud. Models like GLM 5.2 and Kimi K2.7 Code are listed, with Kimi K3 added later.
3. A remote model through LM Link — computation happens on another one of your devices, while the interface stays on the current one.

What's claimed about privacy

LM Studio promises Zero Data Retention for all Bionic users and says it won't train on their data; cloud requests, per the description, are processed in transit and not stored afterward. For local models, the question doesn't come up at all — data never leaves the machine.

These are reasonable terms, but worth keeping two limits in mind. First: the Zero Data Retention promise is a vendor policy, not a technical impossibility of data transfer. You can't verify it from outside, unlike local mode, where there's simply no traffic. Second: web search and any external services stay external no matter where the model computes. More on this logic in the article on whether you can upload work documents to an AI.

What it costs

Local work is free: models through llama.cpp and MLX, offline speech transcription, web search with no data retention, and LM Link across several devices. You pay for the cloud — access to frontier models in Secure Cloud is billed by the token and needs an account with billing set up. For Kimi K3, for instance, the announced rate is $3 per million input tokens and $15 per million output tokens.

Worth pausing here, because this isn't a pricing detail, it's a fork in the road. The "free, no limits" setup works right up until a task outgrows home hardware. From there, Bionic offers you the vendor's own checkout: your own account, your own billing, paying in dollars per token. There's no way around it inside Bionic — its cloud is its own.

Who it fits

Bionic makes sense if you already use LM Studio, keep a set of local models, and are comfortable dealing with engines and loading settings. Then the agentic mode is a natural extension of what's already installed.

Two cases where it doesn't fit. First: work needs to stay inside the perimeter with no exceptions — then the cloud mode is useless, and its presence in the interface mostly creates a risk of accidentally switching a session. Second: paying the vendor for tokens isn't an option, and the local model isn't enough for the task — here Bionic hits a dead end, because there's no alternative to its own cloud inside the app.

Worth factoring in age too: the app is two weeks old. Rough edges in the interface and a shifting feature set are normal for that stage, so it's early to move a critical process into it.

Bionic or Doka

Feature-for-feature, the products are close: both are desktop agents, both work with code and documents, both can compute the model locally. The difference isn't in the list, it's in where free and private stops.

Bionic's strength is switching between a local model, a second device, and Secure Cloud. But its cloud is its own and paid, and Zero Data Retention stays a vendor promise, not a technical impossibility of transfer. You can't verify it from outside — unlike local mode, where there's simply no traffic.

Doka comes at this from the other direction. A local model here isn't one of three options, it's the default mode: the GGUF file downloads inside the app, no need to get familiar with engines or loading settings. Scenarios are built around files, the terminal, schedules, and MCP servers. If a one-off heavy task needs a cloud model, it connects with your own key to the provider you pick, not through the app developer's billing.

Bottom line, no false modesty: if you like LM Studio's cloud and are willing to pay for it, Bionic is a solid, complete product. If you need an agent that stays free and local even when the task gets harder, it makes more sense to start with Doka.

How to compare honestly

Take one folder and one task with a checkable result: go through three documents and put together a summary, or fix a failing test. Run it in both apps on comparable local models, no cloud involved. Look not at how nice the answer reads, but at how clear the confirmations are, the quality of the diff, and how many attempts it takes to get the result you wanted.

Download Doka for free. Other open agents for comparison are covered in the roundups of Goose and Hermes Agent Desktop.