What is new?
On 23 October 2026, Anthropic announced support for MCP tasks and elicitations in Claude through its ClaudeDevs channel. According to the announcement, tasks let Claude hand long-running jobs, such as deep research, to plugins and report back when they are done.
The Claude Code documentation describes the process in more detail. An MCP task is a long job, such as a report export or a build, that an MCP server runs itself using the MCP tasks extension. Claude's tool call returns once the server has accepted the job, and Claude keeps working. When the job ends, a note appears in the transcript and Claude gets the result.
Where do MCP tasks work?
MCP tasks are a public beta in the Claude Code CLI and require version 2.1.304 or later. According to the documentation, all of the following conditions must hold for a tool call to become a task. Otherwise it stays a plain tool call.
- The server supports tasks and is an HTTP, stdio or claude.ai connector server on protocol revision 2026-07-28
- The session fetches feature flags, so not through Amazon Bedrock or other third-party providers and not with telemetry turned off
- Terminal CLI, Agent SDK or GitHub Actions; not the VS Code extension and not the desktop app
- The call comes from the main conversation, not from a subagent; a one-shot claude -p run does not get tasks
What happens while a task runs?
While the job runs, the footer shows a count such as 1 MCP task, and Claude Code checks the status with the server. In /tasks, you can select a task and stop it with x. Claude Code then asks the server to cancel, but the server may still finish the job. Esc does not stop a background task. By default, Claude Code sets no time limit of its own on a background task.
A task on a remote server keeps running after you exit Claude Code. To follow it again, resume the same session interactively, for example with claude --resume, and keep the server's name and connection settings unchanged. A new session or a fork does not follow the task. With a stdio server, by contrast, its process ends together with Claude Code. If Claude Code reports outcome unknown, a second call may repeat the work; first find out from the server what has already been done.
How do MCP servers ask questions?
With elicitation, an MCP server requests structured input in the middle of its work when it is missing information. Claude Code then shows a dialog and passes the answer back to the server. In form mode, the server defines the fields. In URL mode, Claude Code asks whether to open a link in the browser, for example for a sign-in. According to the documentation, no configuration is needed for this.
A running task can ask questions this way as well. Declining or pressing Esc counts as an answer, and the server decides whether the task continues. In non-interactive sessions, Claude Code asks the server to cancel a task that requests input. Since version 2.1.310 of 23 October 2026, when a call times out while an elicitation is unanswered, Claude is also told that no answer came and that it should not repeat the call.
What do MCP server developers need to know?
A server may answer a long tools/call with a task only when the request carries the tasks extension in the client capabilities; otherwise Claude gets a tool error. After that, Claude Code checks the status regularly with tasks/get. On HTTP and connector servers, the task answer has to start within the per-request time limit, so accept the job and answer first, then do the work.
- resultType set to task, plus taskId, status, createdAt, lastUpdatedAt and ttlMs as required fields
- taskId with at most 128 printable ASCII characters and no spaces, answerable across new connections
- pollIntervalMs optional, clamped to 250 to 30,000 milliseconds, 2,000 when absent
- Status input_required with inputRequests; elicitation/create is served only in interactive terminal sessions
What should teams do now?
Check which of your MCP servers run long jobs and whether they support the tasks extension. Test the beta first with a read-only server, such as a report export, in the terminal CLI from version 2.1.304. Because tasks on remote servers can keep running after you exit, it should be clear who stops running jobs and how costs and side effects are controlled.
When a form or link asks for input, always check which server is asking before you enter data or open a link. Record in your AI policy which servers may ask questions. The free AI System Check helps organize tools, approvals and responsibilities across your team. These notes are suggestions, not customer results.
Keep it verifiable
