Claude Code can now run long jobs and stop to ask your permission

Claude Code can now hand off a slow job and keep working, and it can stop partway through to ask you a question. Anthropic shipped both features on October 23, 2026. They work in the Claude Code command-line tool, which is the text-based version developers use. Most everyday Claude users cannot use them yet.

What changed

The features build on MCP, short for Model Context Protocol, a common way for AI tools to connect to other software. The first feature is called Tasks. A tool can start a long job, such as deep research, a build or a report, and give Claude a ticket instead of an answer. Claude checks back on a set schedule until the work is done. Claude Code does not drop a job just because its time limit passed.

The second feature is called Elicitations. A tool can pause mid-job and ask you for something it cannot get on its own, such as confirming a refund, filling in a missing detail or signing in through a browser. Claude Code shows the prompt automatically, with no setup.

Who can use it

Tasks are in public beta. The documentation sets several conditions:

• Claude Code version 2.1.304 or later.
• Terminal sessions, Agent SDK sessions or GitHub Actions runs.
• A tool that speaks MCP protocol revision 2026-07-28.
• A call made from the main conversation, not a subagent.

The VS Code extension and the desktop app do not have Tasks. Third-party providers that do not fetch feature flags, Amazon Bedrock among them, are also left out.

A safety rule worth knowing

Permission prompts only appear in interactive terminal sessions, where a person is present. In an automated run with nobody watching, a job that asks for input is canceled. It does not guess. That keeps an unattended agent from carrying on without a real answer.

What happens next

Anthropic says the features are available now in the Claude Code command-line tool and are rolling out to more Claude products soon. No date has been given. No official newsroom post accompanied the launch, and the material does not mention a price. The record is a developer thread and the documentation. David Soria Parra, an MCP co-creator at Anthropic, previewed the feature in a keynote on October 22.

The MCP specification itself did not change. This is Anthropic building new behavior on top of it. For teams that build AI tools, slow jobs and consent checks no longer need custom workarounds. For everyone else, the practical advice is to wait for the wider rollout.