Claude MCP Tasks: overnight jobs arrive, and so does the double-refund problem
Anthropic added support for MCP Tasks and Elicitations in Claude on Friday evening. Tasks let Claude hand a plugin a long-running job, such as deep research, and get a report back when it's done. Elicitations let a server stop mid-task and ask the user a question. Announcement: https://x.com/ClaudeDevs/status/2113698001889927608
In Claude Code the docs are specific. MCP tasks are in public beta in the CLI and need v2.1.304 or later. A task on a remote server keeps running after you exit, and you pick it up again by resuming the session. Main-conversation MCP calls that run past two minutes move to the background automatically. In a non-interactive session, a task that asks for input gets cancelled instead of hanging. Docs: https://code.claude.com/docs/en/mcp Spec: https://modelcontextprotocol.io/extensions/tasks/overview
The most important lines in the docs are the honest ones. Stopping a task "doesn't guarantee the server stops: it may already have done some of the work, or may still finish it." If the server loses track of a task, the transcript says "outcome unknown," followed by: "A second call may repeat the work. Before calling the tool again, find out from the server what was done." Builders noticed straight away that this is the double-refund, double-deploy, double-email problem, now a protocol feature. The docs also add a way to mark tools that are safe to call twice, and Claude Code v2.1.310, released Saturday, changed the timeout message so Claude is told not to repeat a call when an elicitation went unanswered. https://github.com/anthropics/claude-code/releases/tag/v2.1.310
This is the plumbing for agents that work while you sleep, and it's the right plumbing. But it moves the hard part to MCP server authors. Any tool that touches money, deploys or messages now needs idempotency keys and a status query, because the agent will eventually ask twice. Build that before you ship a long-running tool, not after the first duplicate charge.
← Back to all articles
In Claude Code the docs are specific. MCP tasks are in public beta in the CLI and need v2.1.304 or later. A task on a remote server keeps running after you exit, and you pick it up again by resuming the session. Main-conversation MCP calls that run past two minutes move to the background automatically. In a non-interactive session, a task that asks for input gets cancelled instead of hanging. Docs: https://code.claude.com/docs/en/mcp Spec: https://modelcontextprotocol.io/extensions/tasks/overview
The most important lines in the docs are the honest ones. Stopping a task "doesn't guarantee the server stops: it may already have done some of the work, or may still finish it." If the server loses track of a task, the transcript says "outcome unknown," followed by: "A second call may repeat the work. Before calling the tool again, find out from the server what was done." Builders noticed straight away that this is the double-refund, double-deploy, double-email problem, now a protocol feature. The docs also add a way to mark tools that are safe to call twice, and Claude Code v2.1.310, released Saturday, changed the timeout message so Claude is told not to repeat a call when an elicitation went unanswered. https://github.com/anthropics/claude-code/releases/tag/v2.1.310
This is the plumbing for agents that work while you sleep, and it's the right plumbing. But it moves the hard part to MCP server authors. Any tool that touches money, deploys or messages now needs idempotency keys and a status query, because the agent will eventually ask twice. Build that before you ship a long-running tool, not after the first duplicate charge.
Comments