Your local AI assistant may still use the internet

Why it matters

An application can run its language model on your computer while sending search queries to a remote service. The label local does not describe that whole journey. Before using sensitive material, map what leaves the device, what is stored and which services can observe a request.

Read the evidence

Local Deep Researcher documents local model execution alongside configurable web-search services. Its README lists DuckDuckGo and other search providers. That is enough to reject the assumption that a local model implies an offline research workflow; it is not a packet-level audit of any particular installation.

How to use this resource

Use a harmless sample document and make a boundary table before any live trial. For each component, name its input, destination, storage and deletion route.

• Trace model inference, embeddings, retrieved documents, tool calls and logs separately.
• Inspect configuration for remote endpoints, optional tracing and fallback providers.
• Mark which functions are expected to fail when network access is removed.
• If a test is authorized, compare the observed requests with your table and record surprises.

Limits and evidence

The baseline is CooScope’s 2 October 2026 documentation review; the public repository was rechecked on 4 October. No software was installed, network traffic captured or privacy guarantee established. A future test must use an agreed environment and non-sensitive inputs.

Put it to work.

• Map every external destination.
• Check optional integrations, not only defaults.
• Use non-sensitive test inputs.
• Keep documented and observed behavior separate.

Go to the source.

Original material, with its own scope and publication date. This note is an explanation, not a CooScope performance test.

1. LangChain — Local Deep Researcher repository and configurationLink / source review: 2026-10-04