Your AI Can Finally Order Errands—and the Implications Go Far Beyond That
On May 26, 2026, Meituan shipped a seemingly minor update: the Errand Skill. Without writing a line of code, users install it and simply tell their AI assistant, "Pick up my package from the station," and a courier shows up.
On the surface this looks like replacing app-based ordering with voice. But the more fundamental shift is this: for the first time, AI connects to the physical world through human executors. Large models have lived in the logical world—reasoning, coding, calling APIs—yet unable to touch reality. Here, AI thinks, plans, and decides, while human couriers execute the actions. It's not automation; it's a new mode of Agent–human collaboration.
What Can It Do?
The Skill covers three scenario types:
- Pickup & delivery: take items from point A to B (documents, flowers, cakes).
- Favors: restaurant queuing, hospital registration lines, moving things, trash disposal.
- Buying: vegetables from a market, batteries from a convenience store, cold medicine from a pharmacy.
- Preview first, submit later: preview returns fees without placing the order; submission requires a
--confirmparameter. Preview and submission must occur in the same process or the token becomes invalid. - Amount gating: fees over 100 yuan require extra confirmation—a product-design safeguard against AI misinterpretation causing costly mistakes.
- Address completeness validation: pickup address, drop-off address, and phone number are all mandatory.
- AI as the thinking layer: understanding, planning, deciding.
- Human executors as the physical layer: couriers pick up, carry, deliver.
- The Skill as the connection layer: a standardized interface letting any AI assistant dispatch this human-executor network.
- AI misinterpretation: "buy some cold medicine" could yield the wrong brand or dosage; whether medicine purchases have special restrictions is unclear.
- Privacy: the AI pulls address books and infers frequent addresses; the README states Meituan only provides interfaces and Skill definitions and takes no responsibility for the AI assistant's runtime environment—safety responsibility is dispersed between users and assistant providers.
- Liability: if AI misunderstands an instruction (e.g., wrong delivery address), who is responsible—user, assistant, Meituan, or Skill developer?
- Over-reliance: could frictionless ordering inflate demand and stress couriers at peak times?
- GitHub repo:
meituan/MT-Paotui-For-Client - ClawHub: clawhub.ai/meituan-tech/mt-paotui-for-client
- Meituan official release (2026-05-26)
- Wallstreetcn coverage of the Errand Skill launch
Technically it is a standard OpenClaw Skill—an SKILL.md file plus a set of Node.js scripts. Installation is minimal: send the GitHub link to your AI and ask it to install the Skill. First use redirects to the Meituan app for authorization; afterwards no repeated login is needed. The AI automatically pulls your address book sorted by recency, uses POI search when addresses can't be matched, and auto-detects item categories. Before every order, the AI shows a full preview card (addresses, item, fee, ETA); the order is only submitted after you reply "confirm." Orders over 100 yuan trigger an additional confirmation. After ordering, you can ask the AI for real-time delivery progress.
Three Layers of Design: More Than an API Wrapper
Layer 1: Scene understanding — converting natural language into structured intent. Saying "pick up my package at the station and bring it home" requires identifying the scenario (pickup-and-delivery), extracting entities (pickup address, drop-off address, item), and resolving addresses ("station" matched against the address book, "home" against the default). Item type (package → document class) and weight are also inferred. This is not simple keyword matching.
Layer 2: Risk control — two-step confirmation and amount gating:
This layer shows Meituan understands that AI comprehension can drift, and real-world transactions need human-in-the-loop.
Layer 3: Ecosystem embedding — the most crucial step. Meituan did not lock users into its app; it packaged the capability as a standardized Skill open to the entire AI assistant ecosystem. It's compatible with OpenClaw local and cloud deployments and third-party clients (Cursor, Qclaw, etc.), with code open-sourced on GitHub and ClawHub. Users can order through any connected AI assistant, including WeChat and Feishu integrations. Meituan is turning its fulfillment network into infrastructure anyone can call.
The Real Significance: AI's First Standardized Interface to the Physical World
Past AI Agents had clear boundaries: operate computers, call APIs, process information—all in the digital world. The Errand Skill introduces a new paradigm:
This is not "AI replacing humans" but "AI dispatching humans." AI has no body, yet it can direct embodied people to perform physical actions. Imagined extensions: a housekeeping Skill directing cleaners, a repair Skill diagnosing faults and booking technicians, a procurement Skill ordering stock automatically. Each Skill becomes a "physical-world API," transforming AI from a talking search engine into a commander of real resources.
Competitive Landscape
In April, UU Errand released its own errand Skill, supporting Qoder, Cursor, Claude, and OpenClaw, with quote requests, order placement, order management, and real-time courier tracking. 2026 is being called the year of enterprise-scale multi-agent deployment, as Chinese platforms upgrade AI agents from recommendation tools to execution infrastructure.
Meituan's strategy: open the front-end entrance (any AI assistant can initiate requests), retain core control (payment stays in the Meituan app), and standardize the interface to lower integration barriers. It's not abandoning the app but opening a new traffic entrance: being "chosen by AI."
Concerns and Boundaries
A Starting Point, Not an Endpoint
The Errand Skill is essentially a prototype of a "human-executor API." It demonstrates that standardized interfaces can let any AI assistant dispatch real-world labor; that two-step confirmation balances convenience and safety; and that open ecosystems capture AI-era traffic better than closed apps. Future Agents may not just operate computers but genuinely influence real life—a bridge extending AI from the virtual-logical world into physical reality.