TL;DR: LM Studio Bionic (launched July 16, 2026) is a standalone coding and work agent for open models — and no, it does not silently ship your code to the cloud. Cloud inference is gated behind a signed-in account, billing, and prepaid credits; local and LM Link sessions need no account at all. The two real egress vectors are opt-in cloud models and the agent’s web search tool.

After this guide you’ll be able to:

• Say exactly which Bionic configurations send data off your machine, and lock a session to local-only inference
• Wire LM Studio into Cline and Claude Code correctly — and explain why Bionic itself is not the thing your editor connects to
• Decide when Bionic beats a Cline + Ollama stack for private coding, and when the closed-source trade-off should push you elsewhere

Honest take: Bionic with a local root model is one of the most private coding agents you can run today — the cloud tier literally cannot activate without you setting up billing first. The honest asterisk is that both LM Studio and Bionic are closed source, so “private” rests on documentation and contracts, not auditable code. If zero-trust matters more than polish, Cline + Ollama is still the stack you can read.

LM Studio spent three years as the friendly desktop app for downloading GGUF files. On July 16, 2026 it shipped Bionic, a separate agent app that hit 270 points on Hacker News in a day — and immediately triggered the question this article answers: if the pitch is “the AI agent for open models,” why does it have a cloud tier, and what exactly leaves localhost when you use it for coding?

Our sister site already covered which open models actually run Bionic locally and what hardware you need. This article owns the coding-tool layer: the privacy model as officially documented, and how Bionic does (and doesn’t) fit into a Cline, Claude Code, or Cursor workflow. Everything below was verified August 10, 2026 against LM Studio’s official documentation — read through the lmstudio-ai/docs GitHub repository, which is the source that publishes to lmstudio.ai/docs.

What Bionic actually is (and isn’t)

Bionic is a new desktop app — macOS, Windows, and Linux — separate from LM Studio, built for agentic work rather than chat. It has two project types: Work Projects for documents and research, and Code Projects that connect the agent to a local working directory with file, search, Git, and shell tools. Point it at a repo and it indexes the folder, shows the Git branch, searches the codebase, edits multiple files, runs shell commands, and presents inline diffs for review. It can split large investigations into parallel sub-sessions.

What it is not: a server. Bionic exposes no OpenAI-compatible endpoint for other tools to call. If you came here hoping to type a “Bionic base URL” into Cline or Cursor, that URL does not exist — and we’ll get to what you should use instead.

The engine underneath is the LM Studio runtime (llama.cpp and MLX), so the local model catalog is the same GGUF/MLX ecosystem LM Studio users already know. You download models inside Bionic under Settings → Local Models → Explore, with format filters and device-fit information.

The three places a Bionic session can run

Every session picks its model from one picker, and the choice determines the data path. This table is the whole privacy model:

	Local	Remote (LM Link)	Cloud (Secure Cloud)
Where inference runs	Your machine	Another device you own	LM Studio’s US-based cloud
Does your prompt leave your machine?	No	Yes — end-to-end encrypted to your own device via Tailscale	Yes — to LM Studio’s inference partners
Account required	No	No	Yes, signed in with billing set up
Cost	Free	Free (up to 5 devices on the free tier)	Prepaid credits, pay-as-you-go
Models	Anything that fits your RAM/VRAM	Anything the remote device runs	Frontier open models (large MoE-class)

Three doc-verified facts anchor this. First, from the billing docs: “Cloud models require an LM Studio account with available credits. Local and LM Link models do not use Bionic cloud credits, and can be used without an LM Studio account.” Second, from the credits page: “Credits are only consumed when you select and use cloud models.” Third, cloud models require “a signed-in LM Studio account with billing set up” plus available credits before they will run at all.

That kills the scariest version of the launch-week worry. A fresh Bionic install with no account cannot route your code to the cloud — there is nothing to bill it to. The failure mode to actually watch for is subtler, and it’s next.

The two real egress vectors

Vector one: you signed in, bought credits, and the model picker now has cloud entries. Once billing exists, a cloud model is one dropdown click away from any session, and Bionic sessions inherit a default Root model (Settings → General). If your root model is a cloud one, every new session starts with prompts — and in a Code Project, file contents the agent reads — going to LM Studio’s Secure Cloud. LM Studio’s stated terms are Zero Data Retention with transient processing, and founder Yagil said on the Hacker News launch thread that ZDR was negotiated contractually with the inference providers, not just posted as policy. Take that for what it is: a contractual promise from a closed-source vendor. There’s no published server code or independent audit to check it against — which was the top-voted criticism in the HN thread.

Vector two: the web search tool. Bionic’s free tier includes a zero-data-retention web search tool, and search queries by definition leave your machine even when inference is fully local. An agent researching your task composes those queries itself — from your prompt and project context. For most coding work that’s harmless; for a codebase whose names are sensitive, it’s worth knowing the agent can put fragments of them into a search query.

Locking Bionic to local-only, in four steps

1. Don’t set up billing. No credits, no cloud inference — enforced by LM Studio’s own gating, not by your discipline.
2. If you do have an account, open Settings → General and set the Root model to a downloaded local model, so no session defaults to cloud.
3. Check the model name in the session picker before pasting anything sensitive. Cloud, local, and remote models are listed as distinct types.
4. Treat web search as an egress channel and keep it out of sessions on codebases you wouldn’t name in a Google query.

On hardware: local-only Bionic is only as good as the model your VRAM allows, and the models that make agentic coding pleasant lean toward 24GB cards — an RTX 3090 remains the used-market default for that tier. The runaihome Bionic hardware guide covers the model-to-VRAM matrix in depth. If your desk can’t host that, renting a GPU on RunPod and pointing LM Link at it keeps you on hardware you control — as opposed to credits on hardware you don’t.

Bionic vs. your editor: two different jobs

Here’s the part the launch coverage kept muddling. Bionic competes with Claude Code and Cursor’s agent — it is the agent. It does not compete with LM Studio’s server, which is what Cline, Continue.dev, Aider, and BYOK setups actually connect to. Classic LM Studio still exists precisely for this (“For advanced low-level configuration, you can continue to use LM Studio alongside Bionic,” per the Bionic docs), and its integration docs are blunt: point your tool’s base URL to http://localhost:1234.

So the decision is: use Bionic as your coding agent, or use LM Studio to serve a model to the agent you already have. Here’s how each client actually connects:

Client	Connects to	Base URL / config	Agent features on local models
Bionic	its own runtime	none needed	Full (native)
Cline 4.1.7	LM Studio server	LM Studio provider, http://localhost:1234/v1	Full tool-calling loop, model permitting
Continue.dev	LM Studio server	http://localhost:1234/v1	Chat + autocomplete + agent
Claude Code	LM Studio server	ANTHROPIC_BASE_URL=http://localhost:1234	Full, via Anthropic-compat endpoint
Cursor	❌ not localhost	needs a public HTTPS endpoint	Partial even then (see below)

The one command to run first

Whichever client you pick, start LM Studio’s server and prove it’s alive before touching editor settings:

$ lms server start --port 1234
Starting server on port 1234...
Success! Server is now running on port 1234

$ curl http://127.0.0.1:1234/v1/models
{
  "data": [
    {
      "id": "qwen2.5-coder-32b-instruct",
      "object": "model",
      "owned_by": "organization_owner"
    }
  ],
  "object": "list"
}

If that curl fails, stop — no client config will fix a server that isn’t answering. Our LM Studio connection-error troubleshooter walks every failure branch: ECONNREFUSED, the IPv6 localhost trap, the silent 200-on-404, and JIT unloading.

Cline and Continue.dev

Cline ships a dedicated LM Studio provider — select it, leave the base URL at the default http://localhost:1234/v1 (or set it explicitly if you changed ports), and pick the loaded model. One thing to know if you secured your server: Cline’s LM Studio provider had no API-key field as of the March 2026 feature request (cline/cline#9668), so if you enabled LM Studio’s Require Authentication, use Cline’s OpenAI-compatible provider with a bearer token instead. Full walkthroughs: Cline + LM Studio setup and Continue.dev + LM Studio.

Load your model with generous context either way. LM Studio’s own integration docs recommend more than ~25k context for agentic tools, and Cline’s system prompt alone eats a large share of small windows — the classic symptom is an agent that loses the thread two tool calls in.

Claude Code on a local model

This one is officially documented by LM Studio, which added an Anthropic-compatible POST /v1/messages endpoint specifically so Claude Code can talk to it:

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio
claude --model openai/gpt-oss-20b

And the trick that makes this genuinely useful: with LM Link, localhost:1234 on your laptop can transparently route to a model loaded on your desktop GPU rig — the API surface stays local while inference happens on the other device, end-to-end encrypted. Same environment variables, no cloud in the path.

Cursor: the wrong client for this job

We hit this wall while testing, and it’s worth spelling out because it keeps coming up in setup threads. Cursor’s custom-model support (Override OpenAI Base URL) requires a publicly reachable HTTPS endpoint — it will not call http://localhost:1234, because Cursor’s backend makes the request, not your editor. The workaround is exposing your LM Studio server through a tunnel like ngrok, at which point your “local” privacy story now includes a tunnel provider. Even then, the override only applies to features running on standard chat completions: Tab autocomplete stays on Cursor’s own backend regardless, and community bug reports through mid-2026 document custom base-URL models being silently ignored by Cursor’s subagents. The fix is not a better tunnel. If local models are the point, use a client built for them — Bionic, Cline, or Claude Code — and let Cursor be a cloud tool. Our Cursor + Ollama guide covers the tunnel route if you insist.

So who should actually use Bionic for coding?

Use Bionic if you want a private-by-default coding agent without assembling a stack. The gating is real: no billing, no cloud, verified in the official docs rather than promised in a blog post. Diff review, repo indexing, and parallel sub-sessions come free, and the same app handles document work.

Stay on Cline + a local server if you need your privacy story auditable end to end. Cline is open source; Bionic and LM Studio are not, and the HN launch thread was right to hammer that point — a proprietary harness for open models is a genuine tension, and “Zero Data Retention” is a contract term, not a property you can verify. Our privacy-first Cline setup is the equivalent stack with readable source.

Skip both and wait if you were hoping Bionic would be a drop-in backend for your existing editor. It isn’t one, and LM Studio clearly intends classic LM Studio’s server to keep that job.

FAQ

Is LM Studio Bionic free? The app and all local-model usage are free, including LM Link connectivity for up to five devices. Only Secure Cloud inference costs money, via prepaid credits; a “Bionic Pass” subscription was listed as coming soon as of late July 2026.

Does Bionic send my code to the cloud by default? No. Cloud models require a signed-in account with billing set up and available credits, per LM Studio’s official billing docs. Without those, only local and LM Link inference is possible. The web search tool is the one channel that reaches the internet on the free tier.

Can Cline or Cursor use Bionic as a model backend? No — Bionic exposes no API endpoint. Editor integrations go through classic LM Studio’s server on localhost:1234, which offers both OpenAI-compatible and Anthropic-compatible endpoints.

Does Bionic replace LM Studio? It’s a separate app, and LM Studio’s docs explicitly position the two side by side: Bionic for agentic work, LM Studio for low-level configuration and serving.

Is LM Link actually private? LM Link is built in partnership with Tailscale, and LM Studio documents all inter-device communication as end-to-end encrypted. Your prompts travel between your own devices; nothing is documented as passing through LM Studio’s inference cloud.

Sources

• Introducing LM Studio Bionic — Official LM Studio blog
• Welcome to LM Studio Bionic — Official docs
• Choose a Cloud, Local, or Remote Model — Official Bionic docs
• Set Up Billing and Add Credits — Official Bionic docs
• Understand Credits and Usage — Official Bionic docs
• Code Project — Official Bionic docs
• LM Link — Official LM Studio docs
• Integrations: Claude Code — Official LM Studio docs
• Bionic Pricing — Official LM Studio pricing page
• lmstudio-ai/docs — GitHub source of the official documentation
• Add API key field to LM Studio provider — cline/cline issue #9668
• Custom models (OpenAI base URL override) not usable in subagents — Cursor community forum
• LM Studio Bionic hardware guide — RunAIHome

Last updated August 10, 2026. Official LM Studio documentation was read via the lmstudio-ai/docs GitHub repository, which publishes to lmstudio.ai/docs. Pricing, credit terms, and feature gating change frequently; verify current state on the official pages before relying on them.