Long-running jobs and mid-task questions reach Claude Code
Anthropic’s ClaudeDevs account said on October 23 that Claude is getting support for MCP Tasks and Elicitations, available that day in the Claude Code CLI and coming to more Claude products later. The MCP reference in Anthropic’s Claude Code docs, which the announcement links for server authors, narrows the tasks half considerably. Tasks are a public beta that needs Claude Code v2.1.304 or later, a server on MCP revision 2026-07-28, and one of three kinds of session.
Tasks let Claude hand a plugin a long-running job, with deep research as Anthropic’s example, and pick up the result when the plugin reports back. Elicitations let a plugin pause mid-task to ask the user something, such as confirming a refund or filling in a missing detail. Plugins run on MCP, the open source standard Claude Code uses to reach external tools, databases and APIs. The docs now include a section on MCP tasks for server authors and one on responding to elicitation requests.
How a task runs
Tasks first appeared in the 2025-11-25 revision of the MCP specification as experimental support for durable requests, with polling and deferred result retrieval. The 2026-07-28 revision moved them out of the core protocol into an official extension, io.modelcontextprotocol/tasks, and redesigned them. Polling with tasks/get replaces the blocking tasks/result method, a new tasks/update carries input from client to server, tasks/list is gone, and servers can return a task handle unsolicited, without the client opting in per request, according to the 2026-07-28 changelog. The same revision also dropped protocol-level sessions and the initialize handshake. Every request now carries its protocol version and client capabilities in _meta.
Claude Code implements the extension version. On a 2026-07-28 connection, each request carries Claude Code’s client capabilities at params._meta["io.modelcontextprotocol/clientCapabilities"], with "io.modelcontextprotocol/tasks": {} under extensions. That is what a server checks before handing back a task.
Once a server returns a task, Claude Code polls tasks/get every 2,000 milliseconds, unless the server sets pollIntervalMs, which Claude Code clamps to between 250 and 30,000. The field is optional, but it can’t be null. A taskId can be at most 128 printable ASCII characters, with no spaces. Claude Code doesn’t use tasks/result, tasks/list or notifications/tasks, so the docs tell authors to put the outcome in their tasks/get responses.
A running task can also ask for things. Its inputRequests may hold elicitation/create, which Claude Code serves only in interactive terminal sessions, or roots/list, which it serves everywhere. Keys beyond 16 per task are ignored. When Claude Code can’t serve a request, it asks the server to cancel the task.
Which sessions get tasks?
The docs list the conditions a tool call has to meet before it becomes an MCP task:
- the server is an HTTP, stdio or claude.ai connector server connected on protocol revision 2026-07-28;
- the session fetches feature flags, which rules out third-party providers such as Amazon Bedrock and any setup with telemetry or flag fetching turned off;
- Anthropic has turned MCP tasks on for that kind of session, which on v2.1.304 or later covers terminal CLI sessions, Agent SDK sessions and GitHub Actions runs, while the VS Code extension and the desktop app don’t have MCP tasks;
- the call comes from the main conversation, not from a subagent.
A one-shot claude -p "<prompt>" run doesn’t get MCP tasks either. So in practice, “available today in the Claude Code CLI” means interactive and SDK-driven sessions on Anthropic’s own API, with the server already speaking the new revision. The tweet says Tasks and Elicitations are rolling out to more Claude products soon but gives no date.
Where the refund example stops
In the CLI, elicitation shows up as an interactive dialog, either form fields or a browser URL, and the Elicitation and ElicitationResult hooks can intercept and override a response before it goes back to the server. URL mode was itself added to the spec in the 2025-11-25 revision, along with default values for primitive types in elicitation schemas.
Anthropic’s refund example needs someone at the terminal. In a non-interactive session, Claude Code asks the server to cancel any task that requests input. It doesn’t run your Elicitation hooks or call the Agent SDK’s onElicitation callback first. A GitHub Actions run, which does get tasks, therefore can’t answer a task’s mid-job question: the task is cancelled.
Our analysis
What we checked
On October 24 we read the tasks and channels sections of the Claude Code MCP docs, the Claude Code changelog and npm publish times, and searched the shipped 2.1.310 Linux x64 binary for the protocol identifiers.
The announcement presents both features as arriving now. In the CLI, elicitation has existed for seven months, and tasks shipped four days before the post as a beta with no changelog entry of its own. The binary matches what the docs say about the extension, so a server built to the docs should meet the client it expects, within the session limits above.
| Feature | Claude Code version | Published to npm |
|---|---|---|
| MCP elicitation (form or URL dialog, hooks) | 2.1.76 | March 14 |
| URL-mode elicitation on 2026-07-28 connections | 2.1.281 | September 23 |
| MCP tasks, public beta | 2.1.304 | October 19 |
- No changelog line through 2.1.310 announces MCP tasks. The only related entry fixes servers that negotiate 2026-07-28 but omit
resultType,ttlMsorcacheScope, which had left them showing no tools or failing every call. - The 2.1.310 binary contains
io.modelcontextprotocol/tasks, the five statusesworking,input_required,completed,failedandcancelled, and a 128-character printable-ASCII task-ID check, all consistent with the docs. - It also contains a bundled MCP SDK client with
tasks/resultandtasks/listmethods. That shows only that the strings are present in minified code. It doesn’t show Claude Code calling them, and the docs say it doesn’t. - The spec’s versioning page lists 2026-07-28 as the current protocol version. It isn’t a draft.
- The session limits and the cancel-on-input behaviour come from the docs. We didn’t test them against a live server.
If your server also pushes events
Channel servers push Telegram, Discord or webhook messages into a session so Claude can react while you’re away, and they fall on the other side of the version line from tasks. On the v2 runtime, a channel server that negotiates revision 2026-07-28 can’t deliver channel messages, so Claude Code doesn’t register it as a channel. A channel server that doesn’t support that revision connects on the earlier handshake and registers as before. The same revision is required for tasks, and Claude Code asks stdio servers for it by default.
Up next
OpenAI’s agents broke its sandboxes before training resumed
OpenAI resumed frontier RL training on October 1, 11 days after one of its models reached the live internet through a misconfigured DNS resolver. Do your own agents run with live internet access, or fully cut off?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

