Anthropic shipped support for MCP Tasks and Elicitations in the Claude Code CLI on October 23, 2026 — the client-side move that turns Model Context Protocol (MCP) tools from single request/response calls into durable, interruptible work. David Soria Parra, the MCP co-creator at Anthropic, previewed the feature in his keynote at AGNTCon & MCPCon North America on October 22; the shipping record is the official developer thread the following day. The distinction that matters: this is a client-feature act on top of the spec layer, not a change to the protocol specification itself.
MCP Tasks let a server hand Claude a long-running job — deep research, a build, a report — and report back when it is done. The tool call returns a task handle instead of a result: a task id, a status such as working or input_required, timestamps, a TTL that may be null, and a polling interval the client honors by calling tasks/get until completion delivers the normal tool result. Claude Code does not drop a task merely because its TTL has passed, so a durable job survives its own deadline. The practical effect is that the agent stops blocking on slow tools; the server owns the clock.
Elicitations are the other half of the pattern: a structured way for a server to pause mid-task and ask the user something it cannot obtain itself — confirming a refund, filling in a missing detail, or handing off to a browser for sign-in. Claude Code renders the dialog automatically with no client configuration, in form mode or URL mode. The boundary is explicit in the docs: elicitation is served only in interactive terminal sessions, and a non-interactive session cancels a task that requests input rather than guessing.
The eligibility line is narrow and the docs state it plainly: MCP tasks are in public beta in the Claude Code CLI and require Claude Code v2.1.304 or later, covering terminal CLI sessions, Agent SDK sessions and GitHub Actions runs. The exclusions are just as explicit. The VS Code extension and the desktop app do not have MCP tasks. Third-party providers that do not fetch feature flags — Amazon Bedrock among them — are out. The server must speak protocol revision 2026-07-28, the stateless revision, and the call must come from the main conversation, not a subagent.
Anthropic’s own framing — available today in the Claude Code CLI, rolling out to more Claude products soon — is roadmap, not shipped. No official newsroom post accompanied the feature; the primary record is the developer thread and the documentation, with the October 24 write-ups from secondary outlets forming the coverage wave rather than a second act. That shipped-versus-announced discipline is the metric our AGNTCon Ledger set out to enforce earlier this week.
For anyone building on the agent stack, the change is concrete: long-running server-side work no longer has to be faked with custom polling glue in every plugin, and mid-task human consent no longer has to be routed around. Two failure modes — blocking on slow tools and silently proceeding without input — now have protocol-shaped answers inside one client. The governance backdrop is the same consolidation story covered in the AAIF four-labs analysis: the spec machinery sits under neutral governance while the behavior layer ships from a single vendor’s client.
What this is not is equally clear. The MCP specification did not change on October 23; the July 28 stateless revision remains the spec ground Forkast has already covered, and this feature is Anthropic implementing workflow primitives on top of it. Until the rollout reaches the surfaces most people actually use, the honest reading is narrow: durable tasks and mid-task consent are live for developer-run agents on three surfaces, and everything wider is a promise with a date missing.
