Claude Code can now hand long jobs to MCP servers as tasks and keep working

Anthropic’s developer account announced on 23 October that MCP Tasks and Elicitations are in public beta. The Claude Code MCP docs spell out what that means: “MCP tasks are in public beta in the Claude Code CLI and require Claude Code v2.1.304 or later.”

An MCP task is a long job, such as a report export or a build, that the MCP server runs itself. Claude’s tool call returns as soon as the server accepts the job, so Claude keeps working instead of waiting. The footer shows a count such as “1 MCP task,” and Claude Code polls the server for status. When the job ends, a line in the transcript says so, and the result goes to Claude rather than into the transcript. If Claude is idle when it arrives, it picks the result up on its own.

Tasks that ask questions

A running task can open an elicitation dialog to ask you for input, which is the second half of the announcement. If /tasks is open, the dialog stays hidden until you close it. Declining or pressing Esc is itself an answer, and the server decides whether to carry on. In a non-interactive session, Claude Code asks the server to cancel any task that wants input. We explained how MCP elicitation works earlier this month.

Stopping, exiting and resuming

• Stop a task in /tasks by selecting it and pressing x, or ask Claude. That sends a cancel request, though the server may still finish. Esc does not stop a background task.
• A task on a remote server keeps running after you quit, and no exit option cancels it. Resume the same session interactively (claude --resume, --continue or /resume) and Claude Code picks the task up again if the server still knows it.
• A stdio server’s process stops when Claude Code exits, so unless that server says its work outlives the process, finish or stop its tasks first.
• A new session or --fork-session does not follow the old task, and neither does a one-shot claude -p run.

Where it works

All of these must hold for a call to become a task:

• The server supports tasks and connects over HTTP, stdio or as a claude.ai connector on MCP protocol revision 2026-07-28. Forcing the older SDK generation or legacy negotiation rules it out.
• The session fetches feature flags, which excludes third-party providers such as Amazon Bedrock and setups with telemetry or flag fetching turned off.
• The session type is enabled. On v2.1.304 or later that is terminal CLI sessions, Agent SDK sessions and GitHub Actions runs. The VS Code extension and the desktop app do not have MCP tasks.
• The call comes from the main conversation, not a subagent.

Agent SDK sessions and claude -p runs with --input-format stream-json also need CLAUDE_AUTO_BACKGROUND_TASKS=1, and they must keep their input open, because the run does not wait for tasks after input closes.

Practical takeaway

Update Claude Code to 2.1.304 or later and use the terminal CLI if you want this today. If you maintain an MCP server with slow tools, implementing the tasks extension in protocol revision 2026-07-28 lets Claude Code keep working instead of blocking on your tool, and lets the job ask the user for what it needs.