Claude Code MCP Tasks + Elicitations: Coding Agents Get Background Jobs and Human Forms
Two structural limits have defined coding agents since the start: long jobs block the session, and mid-task decisions mean either the agent guesses or you babysit it. The Claude Code 2.1.304 public beta—announced October 23, 2026, with a practical writeup from Upstash—addresses both. MCP servers can now expose background Tasks and request structured human input through Elicitations.
MCP Tasks: long work stops blocking
Today, when a coding agent kicks off a full test suite, a dependency upgrade, or a twenty-minute migration, one of two things happens: the session waits, or you lose track of the job entirely. MCP Tasks make the long operation a first-class object. Your MCP server declares a task; Claude Code runs it in the background; the interactive session keeps moving; completion and results flow back into the conversation when ready.
The mental model is a job queue you already know from CI— except the thing that enqueues, monitors, and reacts to the job is your coding agent, inside the same terminal where it writes the code that the job tests.
Elicitations: forms for decisions only humans should make
Elicitations invert the usual direction of MCP communication. Instead of the agent calling a tool, the MCP server asks the human a structured question—rendered as a form in the CLI: choose one of these migration strategies, confirm this production command, supply the required parameter the agent cannot infer.
This fills the gap between free-text chat (where the agent might misread your answer) and hard tool calls (which fail outright on missing input). It also gives auditability: a structured form response is a cleaner record of what a human approved than a paragraph of chat. Teams thinking about approval workflows at scale should read our OpenAI Decisions API explainer—agent-vendor approval primitives are converging fast.
What to build with the pair
- Verified upgrades: the agent starts a background task that upgrades a dependency, runs the full suite, and elicits your approval only if tests fail in ambiguous ways.
- Batch document pipelines: point Claude Code at a folder; each file becomes a task; an elicitation asks how to handle the ones that fail validation.
- Staged deployments: the agent runs deploy → smoke test → promote as chained tasks, with an elicitation gate before the promote step.
- Research-then-build: a crawl task gathers context in the background while you keep pair-programming; results land when the agent needs them.
For document-heavy pipelines, the OSS office stack in our GenOffice guide and DeepCell's MCP/CLI toolchain are natural task sources.
The surrounding loop: observe, cap, approve, sandbox
Background tasks and human forms only pay off inside a healthy control loop. Pieces AIXList tracks that complete it:
- Monish — pushes agent progress and approval prompts to your iPhone lock screen, so an elicitation finds you even when you are away from the terminal.
- AgentValley — an OSS macOS notch HUD showing Claude Code sessions and sub-agents, model, and status at a glance.
- Cobalt Cockpit — a Claude Code plugin with a HUD and verification gates for what the agent just did.
- Peras — a local MIT-licensed guard that pauses Claude Code or Codex when weekly or per-prompt usage thresholds hit; background tasks make such caps more important, not less.
- kosa8 — run risky tasks inside a microVM with checkpoint/rewind so a bad background job costs a second, not an afternoon.
Adoption checklist
- Update to Claude Code CLI 2.1.304+ and read the official MCP docs at code.claude.com before relying on beta behavior.
- Start with read-only background tasks (test runs, crawls) before anything that writes.
- Design elicitations to be answerable in one glance— enumerated options beat free text for approval flows.
- Give every task a timeout and a dead-man switch; a stalled background job should surface, not rot.
- Log which human approved what; elicitations are your audit trail if a task goes wrong.
MCP Tasks + Elicitations FAQ
What are Claude Code MCP Tasks?
MCP Tasks, in the Claude Code CLI 2.1.304 public beta, let an MCP server expose tasks that Claude Code runs in the background. Long operations—test suites, builds, research crawls—stop blocking your interactive session; the agent starts the task, keeps working, and reports back when it finishes.
What are Elicitations in Claude Code?
Elicitations are MCP-server-initiated form requests. When the agent needs structured human input mid-task—a choice between options, a required parameter, a confirmation—your MCP server can render a form in the terminal instead of the agent guessing or stopping cold.
Which Claude Code version do I need for Tasks and Elicitations?
CLI 2.1.304 or later—the features shipped as a public beta announced October 23, 2026. Run claude --version to check, and expect iteration as the beta progresses.
How are Elicitations different from tool inputs?
Tool inputs are arguments the model fills in when it calls a tool. Elicitations are requested by the server after a flow has started, aimed at a human, and structured—closer to a form in a classic wizard than to a chat question. They exist for decisions only a person should make.
What should I run as a background MCP Task?
Anything that takes minutes and produces a checkable artifact: full test and lint suites, dependency upgrades with verification, dataset generation, document processing, deployment pipelines. Keep irreversible actions behind explicit approval gates.
AIXList is an independent AI tools directory. This article summarizes the public October 23, 2026 Claude Code 2.1.304 beta announcement and third-party coverage; we are not affiliated with Anthropic or Upstash. Beta features change—treat the official Claude Code documentation as authoritative.