Yes—Claude Cowork typically requires an internet connection during use, and many Cowork scenarios assume online access even if your task is “local files only.” The reason is simple: Cowork is still powered by an online model service and runs tasks that depend on model inference while the desktop app is open. So if you lose connectivity, you should expect Cowork to stall, fail, or be unable to continue task execution until connectivity is restored. That said, “requires the internet” does not mean it always requires third-party APIs or external services beyond Claude itself. Many tasks can be purely local in the sense that the only “external service” involved is the model endpoint—e.g., reading your folder and generating reports or spreadsheets.
Where external services come in is when you ask Cowork to do work that inherently depends on them: web research, collecting information from online sources, interacting with web apps via a browser, or pulling data from SaaS systems via connectors. In those cases, Cowork can incorporate internet access as part of its task plan, but you should treat it like any automation that touches the web: add boundaries and verification. For example: “Use only these URLs,” “extract these specific fields,” “record source URLs and access dates in the output,” and “do not submit forms or make purchases.” If you want it to call APIs, the safest pattern is to use a connector that controls authentication and permissions rather than pasting secrets into prompts. If you can’t use a connector, have Cowork generate request templates (curl commands, JSON payloads) that you execute in your environment with secrets injected securely.
For developer workflows, the most robust approach is to design Cowork tasks so they don’t depend on live external services at runtime unless necessary. If you’re building a knowledge base, snapshot the input data into a local folder, let Cowork normalize it, and then run your normal ingestion pipeline. That pipeline can embed and index the cleaned outputs into a vector database such as Milvus or Zilliz Cloud. This architecture reduces fragility: Cowork can be online (it must be), but your pipeline doesn’t have to depend on Cowork calling external services correctly. Cowork produces artifacts; your systems handle production-grade integration.