On October 6, 2026, Anthropic moves new Claude Cowork tasks on Pro and Max to its servers. The “Only on your computer” setting disappears. Anthropic’s announcement makes the scope explicit.
For a business choosing AI tools, that raises a practical question: who gets to decide where the work happens?
Our answer at Looski is you. The model, the machine doing the work, and the network making the calls should be deliberate choices in your deployment.
What changes, and what stays local
Existing Cowork tasks started locally stay on the computer. Local folders also remain there; cloud tasks can fetch individual files through the open desktop app when needed.
Cloud execution brings useful benefits: work can continue after a laptop closes, and sessions can follow you across devices. Access to files on that laptop still needs the desktop app running.
The announced change concerns Pro and Max. Anthropic still documents local sessions and separate cloud controls for Team and Enterprise.
That scope matters. This is a specific reduction in deployment choice for those plans, rather than evidence that every Claude product is abandoning local execution.
A local app does not necessarily mean a local model
Claude Code retains local execution. Anthropic also explains that it sends prompts and model outputs across the network to interact with the model. Its data-flow documentation separates those two facts.
Running the tools on your computer and running the model on your computer are different arrangements. A desktop icon alone tells you little about where inference happens.
There are three decisions to make:
| Decision | What to establish |
|---|---|
| Where the workflow runs | Which machine executes code, reads files, and performs actions? |
| Where the model runs | Which machine or provider receives the prompt and generates the response? |
| Where calls originate | Which network sends requests to your tools and external services? |
A workflow can combine all three in different ways. You need to know the complete path before deciding whether it fits your business.
Where the call comes from matters too
Anthropic says Cowork connectors reach external services through its cloud. That differs from local browser and computer tools, which can use the desktop machine. See its plugin guide and Team documentation.
For your deployment, establish which machine makes each request, which credentials it uses, and what information it sends. That affects access to internal systems and the network rules your team needs to enforce.
A request sent from your office to an external model still leaves your office. Choosing its origin does not change its destination. Both belong in the design.
With Looski, you choose where the work belongs
Our principle is straightforward: run compatible models where you decide, and make calls from infrastructure you choose. We scope and validate that arrangement with your team.
Looski’s on-premises approach uses compatible open-weight models on hardware in your environment. Inference, retrieval, documents, and operational data are designed to stay on site, with that boundary verified during deployment.
You decide which devices and applications may call the model. You also decide which workflows may contact an external service, what they may send, and which actions need human approval.
Optional external integrations are explicitly scoped. A requirement to keep a workflow local should not quietly turn into permission to route it through a hosted fallback.
For example, a document workflow can extract and prepare information on your own machine, then present the result for staff review. A separate approved workflow can call an external service when that is useful and permitted.
Those are choices to make with your team before activation. Model compatibility, available memory, speed, and integration support determine what we can deliver in the agreed environment.
Keep the decision with the business
Owning the deployment brings responsibilities: capacity planning, updates, backups, maintenance, and access controls. Local compute has costs and limits. Choosing it should be a considered operating decision.
The same is true of cloud services. They can be the right choice when their capabilities and data flows fit the job. The important question is whether your business can keep choosing as its needs change.
We recently wrote about model retirement and business continuity. This is another part of that discussion: a useful workflow depends on decisions about infrastructure as well as model quality.
Your models. Your approved callers. Your chosen infrastructure.
Bring Looski one workflow. We will map where it should run, which systems it may reach, and the controls your team needs before putting it to work.
Frequently asked questions
Which Cowork plans does the October 6 change affect?
The announcement covers new Cowork tasks on Pro and Max from October 6, 2026. Existing local tasks stay local. Team and Enterprise have separate documented controls.
Does running Claude Code locally mean the model runs locally?
No. Local execution describes where the tools run. Anthropic documents network requests to the model, including prompts and model outputs.
Can Looski keep a workflow local while allowing selected external calls?
Yes, where supported by the agreed deployment. We scope the local model, permitted callers, integrations, and outbound requests with your team. External calls require explicit approval; hardware capacity and model compatibility still apply.