This is absolutely incredible.
Claude Code just got two amazing new features that drastically improve how it works with external tools for building software faster.
First there’s the MCP/Plugin Tasks feature — before now Claude Code could never make calls to critical tools without blocking the main thread — and this led to serious wasted time and flow breakdown.
But now this has finally been fixed — Claude Code can now run extremely long tasks fully in the background — and still be able to work on any additional task of your choice.
Demo: Running tests with new Tasks feature — while doing other dev work at the time:
And this still wasn’t enough — they gave us another powerful feature.
Now with Elicitations, all your plugins can now directly interact with you in Claude Code — to get critical input from you — like asking for your personalized input or making sure you approve of sensitive actions.
Demo: Deploy a branch using the new Elicitations feature:
All this combines to save us hours, speed up our development, and give us much greater control over all our coding agents.
1. Asynchronous execution: No more blocking tool calls
Before now, MCP tool calls mostly worked like this:
- Claude sends a request to a tool.
- Claude waits for the tool to finish.
- If the tool takes too long, the request might time out.
MCP Tasks changes this complete.
The server can return a Task ID immediately and run the job in the background. Claude uses that ID to check progress while working on other things.
Real-World Scenario: You ask Claude to run a 15-minute security scan.
Instead of waiting, Claude can review your dependencies or write documentation. Once the scan finishes, Claude picks up the results.
2. Automatic backgrounding for older MCP tools
Your existing MCP servers don’t need to support the new Tasks feature to benefit from backgrounding.
Claude Code can automatically move eligible tool calls into the background after two minutes.
Real-World Scenario: An older MCP server takes five minutes to generate an analytics report.
- After two minutes, Claude backgrounds the call.
- Claude continues working.
- When the tool finishes, Claude gets the result.
You don’t need to rewrite your MCP server.
Older tools still have their usual timeout limits. Native MCP Tasks offer better support for long-running jobs.
3. Interactive elicitations: Claude can ask you for input
Sometimes an AI agent needs information it doesn’t have.
Instead of guessing or failing, MCP Elicitations let the tool ask you directly.
The server can show:
- A form for missing information or approval.
- A URL prompt for things like OAuth sign-ins.
Real-World Scenario: You ask Claude to deploy an application.
The deployment reaches production but needs your approval.
The tool asks you to confirm. You approve and the server continues.
You don’t need to restart the whole process.
4. Safety guardrails for unattended agents
Claude Code also handles tasks that need human input when nobody is available to respond.
If a task requests human input while Claude runs without an interactive terminal:
- Claude doesn’t guess an answer.
- It doesn’t wait forever.
- It asks the server to cancel the task.
Real-World Scenario: A GitHub Actions workflow starts a database migration.
Halfway through, the task asks for approval to delete an important table.
Since nobody can respond, Claude Code requests cancellation instead of approving the deletion.
The server still needs to handle cancellation safely.
5. Durable workflows: Tasks can keep running after you leave
With native MCP Tasks, the Claude Code MCP serve owns the task rather than your local Claude Code session.
This means remote tasks can keep running even if you close your terminal or lose your connection.
Real-World Scenario: You ask Claude to process a large dataset. The job takes two hours.
You close your laptop and return later.
If the server saved the task’s state, you can resume your Claude Code session and check the results.
Native background Tasks also have no default Claude Code execution time limit after acceptance.
But the server still controls task retention and retries.
These updates bring us closer to AI agents that can manage long-running workflows without constant supervision. For those of us building AI agents, this means less waiting and more reliable automation.