Claude Code can now run long MCP jobs and pause to ask you questions

What Anthropic shipped

As YAHOO! reported, Anthropic added support for two MCP features to the Claude Code command-line tool on October 23, 2026. MCP, the Model Context Protocol, is an open standard for connecting AI models to outside tools. The two features are called Tasks and Elicitations.

MCP co-creator David Soria Parra showed the work in his keynote at AGNTCon & MCPCon North America on October 22. The developer thread posted the next day is the official record that it shipped. No newsroom post came with it.

One distinction matters. The protocol's written specification stayed as it was. Anthropic built new behavior into its own client on top of what the spec already describes.

How tasks and elicitations work

Until now, an MCP tool call was mostly a single question and a single answer. The agent asked, waited, and got a result. That works poorly for deep research, a build or a long report.

With tasks, the tool answers right away with a task handle instead of a finished result. The handle carries an id, a status such as working or input_required, timestamps, a time-to-live (TTL, how long the server intends to keep the job) that can be empty, and a suggested interval for checking back. Claude Code follows that interval, asking for updates through a call named tasks/get until the real result arrives. It also keeps a task alive after its TTL runs out.

The result is that the agent no longer sits blocked on a slow tool. The server keeps track of the time.

Elicitations cover the other problem: what to do when a job needs something only a person can supply. The server can pause mid-task and ask you to approve a refund, supply a missing detail, or finish a sign-in in your browser. Claude Code draws the dialog itself, in either a form mode or a URL mode, with no setup on your side.

Where it works and where it does not

The docs draw the line clearly. MCP tasks are a public beta and need Claude Code v2.1.304 or later. They work in three places:

• terminal sessions of the Claude Code CLI
• Agent SDK sessions
• GitHub Actions runs

Several things are left out. The VS Code extension and the desktop app do not support tasks. Third-party providers that do not fetch Anthropic's feature flags are also out, with Amazon Bedrock named as one. The server has to speak protocol revision 2026-07-28, the stateless revision. And the call has to come from the main conversation, not from a subagent.

Elicitations have a limit of their own. They are shown only in interactive terminal sessions. If a task asks for input in a non-interactive session, Claude Code cancels it instead of guessing an answer. That is a sensible default for anything involving money or consent.

Anthropic says the feature will reach more Claude products soon. That is a plan with no date, not something you can use now.

Who this helps

The clearest winners are people building tools on MCP. Until now, anyone with a slow server-side job had to write their own checking code in every plugin. Anyone who needed human approval mid-job had to find a workaround. Both problems now have a standard answer inside one client.

For everyone else, the effect is indirect. A weekly stock check could stop and ask before spending money. A batch job over a set of videos could check in with you instead of running unsupervised. Those are examples of what builders can now design, not products that exist today.

There is also a design burden. A tool that interrupts you is only useful if it asks at the right moments, and nothing in the feature decides that for the builder.

What to keep in mind

The honest reading is a narrow one. Durable jobs and mid-task consent are live for developer-run agents on three surfaces. The desktop app and editor extension most people use are not among them.

It also shows how the pieces of the MCP world are split. The spec sits under shared governance, while this new behavior arrives through one vendor's client. Whether other clients adopt the same pattern is worth watching.