Claude Code adds scheduled cloud tasks on remote machines with MCP access

• Claude Code now supports scheduled cloud tasks that run on a remote machine with access to "your MCPs, repositories, and more," extending Claude Code from interactive sessions into timed automation cloud tasks.
• Anthropic is simultaneously pushing the Claude Agent SDK as a more general way to build agents; one practitioner described it as "the easiest way to build agents," while Anthropic's linked post frames it as tooling for "computer-powered agents" built on Claude Code agent SDK.
• Claude Code also appears to have added an effort control for skills and slash commands, letting developers tune "how long the model thinks before answering" rather than accepting a fixed default effort control.
• Early user reports suggest the remote-access story is still uneven in practice: Simon Willison says Claude Code for web may be limited to one attached private repo per session, and Hamel Husain reports Telegram messages sometimes are "not being received" until he nudges the terminal multi-repo issue Telegram issue.

The new feature is scheduled cloud tasks: jobs kicked off on a remote machine instead of a local terminal. TestingCatalog's summary says those tasks run remotely with access to "your MCPs, repositories, and more," which matters because it suggests Anthropic is wiring the same tool-and-context surface used in live Claude Code sessions into unattended execution cloud tasks.

That changes the operating model more than the UI. Claude Code had already been moving toward being accessible "from anywhere," as Matthew Berman put it from anywhere, but scheduled cloud tasks turn that into actual automation infrastructure: repository-aware work running off the developer laptop, potentially against MCP-connected services and internal tooling MCP access.

Anthropic's broader message is that Claude Code is becoming a substrate for agents, not just a coding assistant. In a thread linking Anthropic's post, trq212 called the Agent SDK "the easiest way to build agents," and the linked article describes it as a toolkit for building "computer-powered agents" on top of Claude Code agent SDK Claude post.

The same thread sketches the design philosophy behind that push: "skills are the abstraction," "your agent should use a file system," and even "bash is all you need" for many workflows agent thread bash thread. That lines up with scheduled cloud tasks running against repos and MCPs: Anthropic seems to be standardizing around agents that can operate through terminal-native tools rather than bespoke app-specific actions.

A smaller but practical control landed alongside that direction. Retweeted release notes from Lydia Hallie say developers can now set effort level in skills and slash commands, which "controls how long the model thinks before answering" effort control second signal. For engineers building repeatable commands, that implies another latency-versus-quality knob inside Claude Code's reusable command layer.

The biggest caveat in the current evidence is scope and session plumbing. Simon Willison says Claude Code for web does not appear to support checking out two private repos at once, because Git operations seem to go through a local proxy that only authenticates the private repo attached to that session multi-repo issue. If accurate, that would constrain cross-repo maintenance and monorepo-adjacent workflows in the web product.

There are also UX reliability reports around remote messaging surfaces. Hamel Husain says a Telegram-connected setup is a "very nice addition" but sometimes Claude Code is "not receiving the message" he sends, and he sometimes has to go back to the terminal and say "hello" to get it unstuck again Telegram issue unstuck report. Another reaction asks for multiple projects and worktrees in the same Telegram bot, which points to the next layer of operator demand once remote access works reliably worktree request.