Local AI privacy

The direct answer

On-device processing runs on this computer. Private-network hosting uses another computer you control and sends data across that network. Offline operation needs no network after setup. Check the destination of inference, document search, extraction, tools, storage and chat history separately. A local interface does not establish any of these boundaries, and cloud-synced folders can move locally processed data elsewhere.

Start with all processing on one computer. Set services to localhost (also called loopback, usually the address 127.0.0.1), which limits connections to that computer. Keep sensitive files outside synced folders. Keep cloud models, web search, and connected tools off until you know what each one sends and why you need it.

Map the workflow before using sensitive material

1. Write down every component. Include the chat app, model, document search, text extraction, scanned-page reading (OCR), connected tools, and storage locations.
2. Identify every provider. Check whether each service runs on this computer, another computer you control, or a company’s servers.
3. Trace every data type. Questions, files, extracted text, search indexes, answers, activity records, crash reports, usage information, and access keys can each go to different places.
4. Check storage and retention. Find chat databases, model files, document indexes, temporary files, backups, sync, exports, and deletion behavior.
5. Check network sharing. A localhost-only service accepts connections from this computer. A service shared over your local network or the internet can accept connections from other devices.
6. Test the real workflow without outbound access. Pre-download required models, turn off remote features, then temporarily block outbound access and repeat the intended task with a harmless document. This checks whether the workflow needs a remote service; it does not establish what the app sends when access is restored.

Optional advanced check: inspect network traffic

Packet inspection can help investigate an unexpected connection. It is optional and requires familiarity with network interfaces and capture permissions.

1. In Wireshark, select the Wi-Fi or Ethernet interface carrying your connection and start a capture without a capture filter. Include VPN or loopback interfaces separately when the workflow uses them.
2. Open a harmless webpage and confirm that both outgoing requests and incoming traffic appear. If they do not, check the selected interface and capture permissions before interpreting results.
3. Close unrelated network apps, note the time, and perform one chat, upload, or indexing operation using a harmless document. Stop the capture and compare the destinations and timing with the services you configured.

Encrypted traffic usually hides document contents, and a capture alone may not identify the responsible process. Missing interfaces or dropped packets can hide activity. No observed traffic is not proof that an app can never send data. See the Wireshark capture setup guide for interface and permission troubleshooting.

Recent changes: LM Studio Bionic and Ollama 0.32

LM Studio Bionic (announced 16 July 2026) is a separate agent app, not a new version of LM Studio. It can run models on your computer, but it can also use cloud models, and the cloud models need an LM Studio account. The regular LM Studio app is the one to use for local-only work.

Ollama 0.32 (released 11 July 2026) changed what the bare ollama command does: it now launches an agent. The release notes show a cloud-tagged model, glm-5.2:cloud, on that start screen. The notes do not say whether the agent picks that model for you. For local-only chat, run ollama run <model> with a model you downloaded. To turn Ollama's cloud features off, set OLLAMA_NO_CLOUD=1 and restart Ollama; the offline guide shows how on each system. Read the Ollama 0.32 article and the LM Studio Bionic article for the details.

Single-machine private setup

For a straightforward confidential-document workflow:

1. Download the model from a known source before introducing sensitive files.
2. Run both the chat model and document-search model on this computer.
3. Use localhost-only settings unless another device needs access.
4. Store source documents, chats, and indexes outside consumer sync folders.
5. Disable remote providers, web search, plugins, and external tools.
6. Import one non-sensitive test document and check that text extraction, document search, and deletion work.
7. Block outbound network access temporarily and repeat the actual workflow.
8. Document the final configuration so an update or new user does not silently change the boundary.

What success looks like: the intended model and document workflow continues to function with outbound traffic unavailable, and you can identify every stored copy of the source, extracted text, index, chat, and export.

Browser interfaces and remote access

A browser chat page can run entirely on your computer. The page can also connect to a model or document-processing service somewhere else. Opening the page successfully does not tell you where those services run.

Keep a personal setup on localhost where practical. Before sharing it with another device, decide who can connect, how they sign in, whether the connection is encrypted, where activity records and uploads are stored, and which connections are allowed. Do not enable wider access to work around a connection error.

Documents, embeddings, and deletion

Deleting the original PDF may not remove extracted text, document-search indexes, thumbnails, temporary files, chat citations, backups, or exports. Find the application's actual storage locations and test deletion on a harmless document. After deleting, search your disk for remnants of the file or related data, such as filenames or keywords extracted from the document. Check the recycle bin or trash, as files may be kept there until permanently removed. Choose deletion methods appropriate to your storage device. Apple does not offer Disk Utility secure erase for SSDs and recommends considering FileVault when you start using the drive; encryption protects stored data but is not proof that every copy has been deleted. Confirm that all related items, such as search indexes or chat citations, are also removed from associated folders. Re-run document and keyword searches to look for accessible remnants, and check application storage, backups, and exports separately. An empty search result does not establish that deleted data is unrecoverable.

For important work, ask for the supporting passage or page reference and read it yourself. This lets you check whether the document supports the answer.

When the local path is the wrong answer

A local setup is a bad choice when the machine cannot run a model capable of the task, the organization cannot maintain the software-and-storage boundary, required collaboration needs controlled central services, or the workflow depends on external tools that cannot be removed. Use a controlled remote system deliberately rather than pretending a hybrid workflow is entirely local.