Basics
What Are MCP Tasks? Long AI Agent Jobs in Claude Code, Explained
MCP Tasks let an AI agent hand a long job to a tool server and keep working until the result is ready. What they are, how they work in Claude Code, and limits.
MCP Tasks are suddenly showing up in AI news because Claude Code now supports them. If you saw the phrase and wondered what it means, the short version is simple: they let an AI agent start a long job, walk away, and come back for the result.
That sounds small. It fixes one of the most annoying things about AI agents today, which is that they freeze while a slow tool runs. Here is what MCP Tasks are, how they work in Claude Code, and who can use them right now.
What are MCP Tasks?
First, a quick refresher. MCP, the Model Context Protocol, is an open standard that lets an AI model use outside tools: a database, a calendar, a build system, a search service. The AI asks a tool to do something, and the tool answers.
Normally that answer has to come back fast. Many apps give a tool call 30 to 60 seconds. If the job takes longer, like exporting a big report or running a full test suite, the call times out or the AI just sits there waiting.
MCP Tasks change the deal. When a job will take a while, the tool server sends back a task ID right away, a bit like the claim ticket you get at a dry cleaner. The server does the work on its own. The AI holds the ticket, keeps doing other things, and asks "is it ready?" until it is.
How do MCP Tasks work?
Under the hood there are only a few moving parts. The official MCP Tasks extension describes them like this:
- Ask. The AI app calls a tool and says it understands tasks.
- Ticket. If the job is slow, the server replies with a task ID, a status, how long it will keep the result, and how often to check back.
- Check. The app polls with a
tasks/getrequest at that pace. - Answer questions. If the server needs something, such as a "yes, go ahead," the task switches to "input required" and the app replies with
tasks/update. - Collect. When the task is done, the result is exactly what the tool would have returned if it had been instant.
Every task is always in one of five states:
| Status | What it means in plain words |
|---|---|
| working | The job is still running. |
| input_required | The server needs an answer from you or the AI before it can continue. |
| completed | Done. The result is ready. |
| failed | Something went wrong. The error is attached. |
| cancelled | Someone stopped it. The server tries to honor this, but it is not guaranteed. |
The last three are final. Once a task reaches one of them, it never changes again.
MCP Tasks in Claude Code
Claude Code is Anthropic’s coding agent. Its documentation now has a section on MCP tasks, and it says they are in public beta in the Claude Code CLI and need version 2.1.304 or later. That version shipped on October 19, 2026, and AlphaSignal and other outlets covered the feature on October 23.
When a task starts, Claude’s tool call returns as soon as the server accepts the job, so Claude keeps working. When the job ends, the transcript says so and Claude gets the result. You can see running tasks in the /tasks list and stop one there by pressing x. Pressing Esc does not stop a background task.
There are some catches:
- Where it works: terminal sessions, the Agent SDK and GitHub Actions runs. Not the VS Code extension or the desktop app, per Anthropic’s docs.
- Which servers: the server must support tasks and connect on MCP protocol revision 2026-07-28.
- Which accounts: sessions on third-party providers such as Amazon Bedrock are left out for now.
- Local servers: a server that Claude Code starts on your own computer stops when you quit, so its job stops too, unless the work runs somewhere that lasts.
MCP Tasks vs background tool calls
Claude Code already had a simpler trick. Any MCP tool call in the main chat that runs longer than two minutes gets moved to the background so you can keep working. The difference is who holds the job.
| Background tool call | MCP Task | |
|---|---|---|
| Who runs the job | The original call, kept open by Claude Code | The tool server, on its own |
| When it starts | After the call runs two minutes | Right away, when the server decides |
| Survives quitting Claude Code | No | A remote server’s job can, and a resumed session can reconnect |
| Can ask you questions midway | Only through a normal dialog | Yes, built into the task |
MCP Tasks and elicitation
Elicitation is the MCP word for "the tool asks the human a question." Claude Code has had it since early 2026. With tasks, the two fit together: a long job, such as processing a refund batch, can pause and pop up a dialog asking you to confirm. Declining or pressing Esc counts as an answer, and the server decides what happens next. In a non-interactive run, Claude Code cancels any task that asks for input, so nothing waits forever.
What this means for you
Unless you build software, you will not type tasks/get anywhere. But the idea matters. AI agents are moving from "answer me now" to "go do this and tell me when it is done." That is the same shift behind Claude dynamic workflows and Anthropic’s background app Claude Duet. MCP is also how apps like Canva plan to bring their own tools into chatbots, starting with its new image model, Canva Craft.
You can already get that kind of help on your phone without any of the plumbing. Agentic AI gives you a team of ready-made expert agents on iPhone and iPad. You hand over one task, up to four agents work on it together, and you get finished work back. There is no account and no login, and the first 3 tasks are free.
Quick answers
What are MCP Tasks? A part of the Model Context Protocol that lets a tool server answer a slow request with a task ID instead of a final result, so the AI can keep working and collect the result later.
Does Claude support MCP Tasks? Yes, in public beta in the Claude Code CLI from version 2.1.304. Anthropic’s docs say the VS Code extension and the desktop app do not have them yet.
Are MCP Tasks the same as Claude Code background tasks? No. A background task is a slow tool call that Claude Code moves aside after two minutes, and it ends when you quit. An MCP Task is run by the server itself and can be resumed.
What are MCP long running tasks used for? Jobs that take minutes or hours: test and build pipelines, big data exports, batch imports, deep research, and steps that wait for a human to approve.
Are MCP Tasks part of the MCP spec? Tasks first appeared as an experimental feature in the November 25, 2025 spec. They now live in an official extension, io.modelcontextprotocol/tasks, which servers and clients must both turn on.
Do I need MCP Tasks to use AI agents on iPhone? No. An agent app such as Agentic AI handles the setup for you.
Sources: Claude Code Docs: MCP · Claude Code changelog · MCP Tasks extension · MCP spec 2025-11-25: Tasks · AlphaSignal · Pasquale Pillitteri
