Claude Code adds long MCP tasks and user questions in public beta

On 23 October 2026, Anthropic added 2 functions to the Claude Code command-line tool: MCP Tasks and Elicitations. MCP is the Model Context Protocol, a standard that connects AI agents to tools. A tool can start long work and report when it ends. It can also stop and put a question to the user.

How it works

A server gives Claude Code a task ID, not a result. Claude Code checks the status of the task until the work is done. The server keeps the clock.

To put a question, the server stops the task. Claude Code shows a dialog in the terminal. This works only in an interactive session. A session that is not interactive cancels the task.

Where it works

It works on 3 surfaces:

• the terminal command-line tool
• Agent SDK sessions
• GitHub Actions runs

VS Code, the desktop app and Amazon Bedrock do not have it.

The server must use protocol revision 2026-07-28. The task must start in the primary conversation, not in a subagent.

What is not known

Anthropic gave no test results. Anthropic also tells that more Claude tools will get the functions, but it gave no date. The protocol did not change.

Insight: One vendor sets how agents wait for people

A neutral group holds the specification of MCP. But these functions are a client step: Anthropic made the task and consent behavior in its tool. This seems to let one client set how agents wait for a person, because the first behavior in use can become the habit. Builders must know that other clients can be different.