Claude Code MCP Tasks: what they are and how to use them
2026-10-25 · 4 min read
MCP Tasks let a connected MCP server take on a long job, like a report export, a build or a deep research run, while Claude keeps working and collects the result once the job is done. You need the Claude Code CLI at version 2.1.304 or later and an MCP server that supports the tasks extension. You don't have to turn anything on.
Anthropic's developer account announced it on Friday. @ClaudeDevs wrote, "We're adding support for MCP Tasks and Elicitations in Claude," and said Tasks let Claude hand long-running jobs to plugins and get a report back. A follow-up from @ClaudeDevs said both are "available today in the Claude Code CLI, rolling out to more Claude products soon," and sent server builders to the Claude Code MCP docs.
The protocol piece is older. Tasks shipped as an extension in the July 28 MCP spec revision, which we wrote up in The protocol connecting AI to your tools just got rebuilt. The news this week is that Claude now uses it.
What Tasks and Elicitations do
Until now, a slow tool call left Claude stuck until the server answered or the call timed out. With Tasks, the server accepts the job, replies right away with a task ID, and does the work on its own side. Claude Code checks the status in the background and hands the result to Claude when the task finishes.
Elicitations run in the other direction. A server can pause partway through and ask you for something it can't get on its own. In form mode, Claude Code opens a dialog with fields the server defines, a username and password for example. In URL mode, it asks whether to open a link in your browser, which servers use for flows like sign-in that finish outside the terminal. A running task can also open one of these dialogs when it needs an answer from you.
Where it works today
It's a public beta, and a narrow one. According to the docs, MCP Tasks work in the Claude Code terminal CLI, the Agent SDK and GitHub Actions, all on v2.1.304 or later. The VS Code extension and the desktop app don't have them yet. Read the limits before you plan anything around this:
| Requirement | Detail |
|---|---|
| Server support | The server must support tasks on protocol revision 2026-07-28, over HTTP, stdio or a claude.ai connector |
| Provider | Not available on third-party providers such as Amazon Bedrock, or with telemetry or feature-flag fetching turned off |
| Caller | Only calls from the main conversation get tasks. Subagent calls don't |
| One-shot runs | A plain claude -p "<prompt>" run doesn't get MCP tasks |
| Exiting | A task on a remote server keeps running after you exit. A stdio server stops when Claude Code exits |
How to try it
- Run
claude --versionand update if you're below 2.1.304. - Connect an MCP server that supports tasks, with
claude mcp addor through a claude.ai connector. - Ask for something slow. If the server replies with a task, the footer shows a count like
1 MCP task. - Type
/tasksto see it. To stop one, select it and pressx, or ask Claude to stop it.Escwon't stop a background task. - When it finishes, the transcript prints a line saying the task completed. The output goes to Claude rather than your screen, so ask Claude for the result.
If you leave, resume the session interactively with claude --resume or /resume to keep following the task. To make Claude wait on every MCP call, set CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS to 0 in your settings.
Elicitation dialogs appear on their own whenever a server asks. In headless runs, with nobody there to answer, Claude Code cancels any task that requests input. The Elicitation hook can answer certain prompts automatically.
For people building MCP servers
Check that the client lists the io.modelcontextprotocol/tasks extension in its capabilities, and only then answer a tool call with a task result carrying a task ID and a status. Claude Code polls tasks/get until the status is completed. On HTTP servers, accept the job and reply first, then do the work. Our guide to Claude Code mods covers the other way to extend the agent.
Our read
The launch is small, but it fixes something real. Exports, reconciliations and research runs take minutes, and the risky steps need a person to approve them. Tasks handle the waiting and Elicitations handle the approvals, and since both live in an open protocol, any connector vendor can add them. Compare Claude auto mode, where Anthropic went the other way and replaced most approval clicks with an automated check.
If you run a small business, work out the workflow before you pick the tool: which jobs run long, who signs off and at what step, and what should happen when nobody is around to answer. New Face Design's free process audit can help you map that out.