Anthropic added MCP Tasks to Claude Code as a public beta. In plain terms, a tool you connect to Claude Code can now take on a long job, such as a report export, a data sync, or a build, hand back a ticket number right away, and keep working on its own server while Claude carries on. The job can also pause partway through and ask you a question in a small form, for example "approve this refund of this amount?"
We read the Claude Code MCP documentation and the MCP Tasks extension overview on 2026-10-24, after the update showed up in X news coverage on 2026-10-23. This post covers what changed, what a business should check, and where it is and is not worth building around.
What is changing
Before this, an MCP tool call was a single request and a single answer. If the tool took ten minutes, the conversation waited or the call timed out.
Now there are two separate behaviors, and the docs are careful to keep them apart:
- Automatic backgrounding. Any MCP call in the main conversation that is still running after two minutes moves to a background task. Claude keeps working and gets the result later. You can see and stop it in `/tasks`. This kind of background call does not survive closing Claude Code.
- MCP tasks run by the server. A server that supports the tasks extension can answer a tool call with a task handle instead of a result. Claude Code then checks the task status every few seconds. The job lives on the server, so a task on a remote server keeps running after you exit, and Claude Code picks it up again when you resume the same session.
On top of that, a running task can ask for input. The docs call this elicitation. In an interactive terminal session you see a dialog. Declining or pressing Esc is also an answer, and the server decides what happens next. In a non-interactive run, Claude Code asks the server to cancel any task that needs input instead of guessing.
Where it works today
According to the docs, MCP tasks need Claude Code v2.1.304 or later and are on for terminal CLI sessions, Agent SDK sessions, and GitHub Actions runs. The VS Code extension and the desktop app do not have them yet. The server has to support tasks and connect on the 2026-07-28 protocol revision, and the session has to fetch Anthropic's feature flags, which rules out third-party providers such as Amazon Bedrock and setups with telemetry turned off. Calls from subagents stay plain tool calls.
The most important line for anyone who has already connected tools: existing MCP servers do not get this automatically. Whoever built the server has to add task support.
What to check before relying on it
- Does your MCP server support tasks? If you use a vendor's server, check its release notes. If you built your own, it will not until someone adds it.
- Who answers the questions? A task that needs approval can only get it from a person in an interactive session. Scheduled or headless runs cancel those tasks, so approvals in unattended jobs need a different route, such as an approval queue your team checks.
- What happens when someone stops a task? Stopping a task in `/tasks` asks the server to cancel, but the docs say the server may already have done some of the work or may still finish it. Any tool that writes data, sends messages, or moves money has to handle a half-finished job safely.
- Session habits. To get a remote task's result after closing Claude Code, you resume the same session. Starting a new session or forking does not follow the task. Teams that clear and restart often will lose track of results unless the server also records them somewhere else.
How a small business would use this
Most owners will never type `/tasks`. The change matters if you, or someone on your team, uses Claude Code with tools connected to your business systems. A few examples where long jobs and a mid-job question fit well:
- Month-end exports. Pull a quarter of invoices from your accounting system, reconcile them against payments, and stop to ask before writing any adjustment.
- Bulk CRM cleanup. Merge duplicate contacts in batches, and pause when two records disagree on something important, such as an email address.
- Refunds and credits. Prepare a refund batch, then ask for a yes or no on each one above a set amount.
- Large document jobs. Generate a set of client reports overnight on a remote server, and pick up the results in the morning by resuming the session.
The common thread is a job that is too long to wait for and too important to finish without a person saying yes at the right moment.
What we would set up around it
If you asked us to build or update an MCP server for this, we would:
- Add task support only to the tools that are actually slow, and keep quick lookups as plain calls.
- Write clear status messages, so `/tasks` shows progress in words your team understands instead of raw IDs.
- Put the approval question inside the task with the specific record and amount, and treat "decline" as a clean stop.
- Make every write step safe to retry, and log what the job did, so a stopped or interrupted task can be checked and finished by hand. Our post on idempotent webhooks for agents covers the same idea.
- Store results outside the chat as well, for example in a shared folder or your CRM, so nobody depends on resuming the right session.
- Match the permission rules to the tool. Our posts on approval controls for MCP tools and running a secure production MCP server go further.
When it is not worth changing anything
If your connected tools all answer in a few seconds, nothing changes for you. The same goes if you use Claude through the desktop app or VS Code, since tasks are not there yet. And if a job needs a person to approve steps while nobody is at the keyboard, a task alone will not solve it. A scheduled workflow with an approval queue, such as the one in our maintenance request automation guide, is often simpler and easier to audit.
It is also a public beta. We would not move a process the business depends on onto it without a fallback.
How this fits implementation work
We build and maintain MCP servers and agent workflows for small and mid-sized businesses. If you have a slow tool that needs a long-job mode or approval questions, see our MCP development and AI agents services. If you are not sure whether any of this applies to you, our free automation audit looks at your current tools and tells you what is worth building. For background on the protocol itself, start with what MCP servers are, and for long-running agent jobs outside Claude Code, see our post on durable harnesses for long-running agents.
Independence and limitations
OrchestriAI is not affiliated with Anthropic, and nothing here is paid placement. This post reflects the Claude Code documentation as read on 2026-10-24. MCP Tasks are in public beta, so details such as which apps support them, the two-minute threshold, and the cancel behavior may change. Check the current docs before you build on them.
