Is LM Studio safe? What stays on your computer, and what Bionic sends to the cloud

The Morse teamPublished November 8, 20269 min read

As of November 2026, LM Studio is safe and private when the model runs on your own computer: its privacy policy says that if you run models locally, "none of your messages, chat histories, and documents are ever transmitted from your system" (LM Studio). Three things do leave: your model searches and downloads, an update check, and, since July 2026, whatever you send to a cloud model or web search in Bionic, LM Studio's new agent app. The rest of the risk sits on your own machine, in the model files you load, the local server you can switch on, and the MCP tools you install.

We read LM Studio's privacy policy (last updated October 19, 2026), its app terms (version of August 23, 2026), its pricing page, 14 pages of its docs, its blog, its bug tracker on GitHub, and the National Vulnerability Database's entries for llama.cpp, the engine LM Studio runs models with, all on November 8, 2026. LM Studio deserves credit for how short its list is. The policy names the only occasions it hears from your copy of the app, and its offline guide sorts every feature into "does not require connectivity" and "requires connectivity."

LM Studio safety at a glance

Our table, built from LM Studio's own pages on November 8, 2026.

Does LM Studio collect your data?

Not your chats, if the model is local. Beyond email you send it, the policy lists three moments when LM Studio receives anything about your use: when you search for or download a model, when the app checks for updates, and when you use a paid cloud feature (LM Studio).

The update check sends your app version and build and your operating system, and your IP address is "included automatically by our CDN provider." LM Studio says it uses this to learn "the software and hardware spread of our user-base" and does not "track your individual behavior." Model searches go to LM Studio as "anonymized search queries," and the downloads themselves come from Hugging Face. The offline guide adds that the Discover tab fetches each model's stats and download options from huggingface.co, and that runtimes, the engines that run each model format, are downloaded separately (LM Studio).

The policy also explains why a data request gets you nowhere. "Because the application does not include telemetry or user-specific tracking, we are unable to fulfill data subject requests such as providing a copy of your data or deleting your information." If LM Studio's owner, Element Labs, is sold or merged, "information may be transferred to a new business."

Is LM Studio private when you use cloud models?

Less than a local model, because your words leave your computer. Bionic, a separate app LM Studio launched on July 16, 2026, can send a chat to "LM Studio Secure Cloud," and its launch post commits to "Zero Data Retention and never training on your data" (LM Studio). Cloud models need an LM Studio account with credits, while local and LM Link models don't (LM Studio).

The privacy policy is a shade softer than the launch post. It says cloud requests are "processed transiently and not stored after the request completes," and that "All parties involved in processing your data operate under Zero Data Retention or substantially equivalent terms" (LM Studio). It doesn't name those parties or say what "substantially equivalent" allows. It also says the processing applies "if and only if you purchase cloud features AND you choose to use the feature(s)," while the pricing page lists "Limited web search" on the free plan. We couldn't tell from either page whether a free search counts as a cloud feature. Either way, a search query leaves your computer.

Connecting Google Workspace puts that data under the same promise: "neither we nor the subprocessors we engage retain this data." LM Studio's reputation as an app that never needs the internet comes from the original app. It still holds for local models, but not for Bionic once you pick a cloud model.

Is LM Studio safe to download?

Yes, from lmstudio.ai. On March 24, 2026, Windows Defender started flagging LM Studio 0.4.7 as a Trojan and deleting its files, and users filed a bug report the same morning (GitHub). That evening a member of LM Studio's team wrote: "We are now confident this was a false positive. We contacted Microsoft who acted quickly to confirm," and added that the team was "auditing our build machine scripts + envs" anyway (GitHub).

You can't check the desktop app's code yourself. LM Studio's terms call its source code "valuable trade secrets" and forbid reverse engineering it (LM Studio). Its command-line tool, lms, its MLX engine, and its TypeScript SDK are MIT-licensed on GitHub (GitHub). Our local AI vs cloud AI guide sets LM Studio's license beside Ollama's, Jan's, and llama.cpp's.

Can a model file harm your computer?

It can, through a bug in the program that reads it. LM Studio runs models with llama.cpp on every platform and with Apple's MLX on Apple Silicon (LM Studio), and it imports outside models as GGUF files (LM Studio). GGUF holds weights and settings, not code, which makes it safer than the Python pickle files Hugging Face warns about, where "dangerous arbitrary code execution attacks" are possible on load (Hugging Face).

GGUF readers still have bugs. The National Vulnerability Database lists llama.cpp flaws where a crafted GGUF file corrupts memory:

• CVE-2024-21825, published February 26, 2024 and rated 8.8, high: "A specially crafted .gguf file can lead to code execution" (NVD).
• CVE-2026-33298, published March 24, 2026 and rated 7.8, high: an integer overflow before llama.cpp build b7824 "allows potential Remote Code Execution (RCE) via memory corruption" (NVD).

LM Studio ships llama.cpp as a runtime you update apart from the app, under Ctrl Shift R (⌘ Shift R on a Mac) (LM Studio). Keep runtimes current, and prefer models from publishers you recognize. The terms put model licenses between you and whoever made the model, and say LM Studio isn't liable for "any failure of the third party products" (LM Studio).

Is LM Studio's local server safe?

On its defaults, yes, since it answers only your own computer. Turn on Serve on Local Network and "the server binds to a non-localhost address," and the docs warn: "Any bind other than 127.0.0.1 exposes the server beyond localhost; we recommend enabling authentication" (LM Studio). Authentication is off until you switch on Require Authentication, which arrived in version 0.4.0 (LM Studio).

MCP servers carry the larger risk, because they give a model tools. LM Studio's own page says "Some MCP servers can run arbitrary code, access your local files, and use your network connection" and "Never install MCPs from untrusted sources" (LM Studio). The server setting Allow calling servers from mcp.json "can be a security risk if you've defined MCP servers that have access to your file system or private data," so LM Studio only lets you turn it on with authentication (LM Studio).

To reach your models from another place, LM Link is the safer route. It runs over Tailscale's mesh VPN, encrypted end to end, and LM Studio says your devices "are never exposed to the public internet" (LM Studio).

Can you use LM Studio at work?

Yes, free, since July 8, 2025, when LM Studio dropped its separate commercial license (LM Studio). The terms add that if you use the app for your employer, your employer becomes the customer, and that LM Studio's liability for all claims is capped at $50 (LM Studio). Our local AI vs cloud AI guide covers the license wording and model licenses.

How Morse runs open models your laptop can't hold

LM Studio keeps you private by keeping the model on your machine, so your machine sets the ceiling. Its system requirements recommend at least 16 GB of RAM (LM Studio), and the models its $20-a-month Bionic+ plan runs in its US-hosted cloud, like Kimi K3 and GLM 5.3 (LM Studio), are there because a laptop can't run them. Past that line you're back to a provider's promise, and your local history is plain JSON on a disk. Morse is a web app, so it runs in your browser with nothing to install, and it reaches big open-weight models with the prompt sealed.

• E2EE, for the prompts you'd otherwise send to Bionic's cloud. Open-weight models like GLM, Kimi, and DeepSeek run in secure enclaves. Before anything leaves your device, your browser checks the enclave's hardware attestation against versions pinned in the app, then encrypts the message so only that enclave can read it. Our servers pass along ciphertext, and only your browser decrypts the reply. Third-party providers run the enclaves, not Morse, and the guarantee holds as long as the enclave hardware and attestation work as designed. E2EE models can't search the web or make images.
• Private, the default, for questions that need a web search. Open-weight models and xAI's Grok. Morse has zero data retention agreements with its inference, search, and other providers, so they keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
• Anonymous, for GPT, Claude, and Gemini, which no local app can download. Our proxy strips your account, email, and IP address and sends the request under Morse's own account. The provider sees the message without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy.

Your Morse history is stored in your browser, encrypted, and locked when you sign out. Unlocking it takes a key that never leaves your device and one our server gives you at sign-in, so a copied browser profile can't be read on its own. Morse doesn't train AI on your chats, sell your data, or show ads, as our privacy policy says. Clearing the browser or deleting your account erases the chats, and we can't recover them, so save anything you need. Our guide to end-to-end encrypted AI chat walks through the enclave check, and is Ollama private covers the other popular local app.

So is LM Studio safe for you?

• Private chats on your own hardware: yes. Download a model, then work offline. On Morse, E2EE keeps the prompt sealed from us for models too big to download.
• Bionic cloud models: a zero-retention promise from LM Studio and providers it doesn't name. On Morse, your browser checks the enclave before it sends anything.
• Model files: stick to publishers you know, and keep runtimes updated.
• The local server: leave Serve on Local Network off, or turn on Require Authentication first.
• Your history: the JSON files are readable by anyone with your computer. On Morse, history is encrypted in your browser.

Send a first message from the Morse home page with no account, or start free with no card. E2EE chats come with Pro, at $19 a month.