Prasenjit Sarkar (@stretchcloud) on X

What it claims

Claude now supports two MCP extensions, Tasks and Elicitations, according to a post by Prasenjit Sarkar. Before this, a tool call got 30 to 60 seconds before the connection timed out. That was too short for CI pipelines, research runs or training jobs. With Tasks, Claude Code, the Agent SDK and GitHub Actions can hand a server a long job and disconnect instead of holding a connection open. The post says AWS helped build the extension. The quoted Anthropic announcement names deep research as an example of a long-running plugin job.

How it works

A server that accepts a task returns a durable handle. The client polls tasks/get at intervals between 250 ms and 30 s. If the client disconnects and reconnects, the same task ID resumes where it stopped.

Elicitations use the same task path. When a task needs a human to approve an action or fill in a field, it moves to input_required. No second connection or custom webhook is needed. Cloudflare's stateless Workers servers use the same input_required and retry pattern.

flowchart TD
  A[Client calls tool] --> B[Server returns task handle]
  B --> C[Client polls tasks/get]
  C --> D{Task state}
  D -->|working| C
  D -->|input_required| E[Human approves or fills field]
  E --> C
  D -->|completed / failed / cancelled| F[Done]
  C -.disconnect, reconnect same ID.-> C

Evidence

The post gives adoption reports, not measurements. Sentry ran on the new spec before it was finalized. Linear built one MCP server that works unchanged across every client. Cloudflare says its Code Mode server has handled billions of tool calls. The post compares Tasks to async patterns that frameworks already build themselves, such as LangGraph interrupts and run polling in OpenAI's Assistants API. In Sarkar's view, the difference is that the pattern now sits in the protocol, so any MCP server gets it without vendor-specific code.

Limits

The post has no code, payload schema, spec version or release date. It says only that the launch came "this week," days after Anthropic donated Agent Skills to the Agentic AI Foundation, which also hosts MCP. Sarkar flags a risk: a task that runs for hours without a human is a task that has to be trusted for hours. Elicitations are the approval gate, and Sarkar says no one is yet measuring whether teams wire them in or skip them because polling is tedious.