ubereats-mcp
You can use AI to search, prepare, review, and place Uber Eats orders through your own Chrome login, with explicit confirmation and safety checks.
Read ordering policy, caps, today’s spend, and login status (
ue_policy).Search stores/dishes/products and inspect menus, including in-store searches (
ue_search,ue_store).Add items with options and quantities to per-store carts, then list carts and draft IDs (
ue_add,ue_cart).Get a trusted checkout snapshot with items, fees, total, ETA, address, payment method, and a single-use confirm token (
ue_review).Place an order only after explicit user approval, with revalidation of total/store/cart and policy caps (
ue_place).Check active orders and delivery progress, especially after placing or when status is unknown (
ue_orders).
Allows AI to interact with Uber Eats using the user's own logged-in Chrome session: search stores and products, view menus, add items to cart with options and quantities, review checkout snapshots, place orders only after explicit confirmation with a confirm_token, and track ongoing delivery orders.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@ubereats-mcp find a good sushi place and add a California roll to my cart"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
ubereats-mcp
简体中文 | 日本語

介绍页:https://saekisui.github.io/ubereats-mcp/
让 AI 在你自己的 Chrome 里帮你点 Uber Eats(有 Uber Eats 的国家都能用,日本站实测最多):搜店 / 搜商品、看菜单、加购、看结算快照,你点头之后才下单,下单后跟配送。
一个 MCP server + 一套 OpenCLI adapter。AI 用的是你 Chrome 里的真实登录态,地址和支付方式只用账号里已经存好的,对话里从不出现卡号或地址。
非官方项目,与 Uber 无关。只在你自己的账号上、花你自己的钱。
它做什么,不做什么
做
搜店、搜菜、搜商品。按商品名搜(ポカリスエット、風邪薬…)会返回超市 / 药店,并附上命中的商品和价格
看菜单:餐厅照常;超市 / 便利店 / 药店的首页只放几排推荐,可以用
query走店内搜索加购:支持选项(多级选项逐级匹配,例如 套餐 → ドリンクM → コカ・コーラ)和数量
结算快照:店名、商品、各项费用、合计、ETA、配送地址、支付方式
下单:只能在 review 之后、拿着 confirm_token 下
查进行中订单的配送进度
不做
不碰支付密码、卡号、地址簿(弹出要编辑地址的框会直接停住)
不过人机验证,不替你预约配送时间
不会自己下单。必须走
ue_review→ 你在对话里明确同意 →ue_place
Related MCP server: jevpilot
安全设计
关卡 | 做法 |
价格从哪来 | 只信 |
防串台 | 两个独立锚点:draft 本体的店 uuid 必须等于购物车列表里的店;接口回的商品必须覆盖购物车里的商品。任一不符直接报错,绝不返回混合了两辆车的快照 |
confirm_token | review 时发,10 分钟有效,只能用一次,绑定 draft。place 前再读一次快照,总价 / 店名 / 购物车变了就拒绝 |
上限 |
|
下单那一刻 | 页面合计必须等于接口合计才点;凑单弹窗只点「スキップ」,里面任何带 ¥ 的商品按钮都不碰;认不出的弹窗一律停下,并把按钮列表带回来 |
状态不明 | 返回 |
国家和界面语言
Uber Eats 网址里的 /jp 管的是界面语言,不是国家:显示哪些店、用什么币种,跟着你账号的配送地址走。所以这套东西不分国家,只分界面语言:
| 界面 | 状态 |
| 英文 | 读的部分(搜店 / 菜单 / 购物车 / 结算快照)按同一套接口走;下单这一步还没人在英文界面上真跑过 |
| 日文 | 实测最多:十几单真单,凑单弹窗、地址确认框、假阴性这些坑都踩过 |
价格不认币种(¥ / $ / NT$ / HK$ / £ / € 都能读,12,34 这种逗号小数也认),「当天」按 policy.json 的 timezone(不写就是电脑的时区)算。
在日本以外的地方第一次用,请这样走一遍:
只读命令都跑一下:
whoami/search/store/cart/reviewopencli ubereats place --draft <draft_id> --dry-run:会走到结算页、做完所有校验就停,不下单第一单真下的时候,人在旁边看着。脚本认不出的弹窗会停下,并把弹窗里的按钮列表带回来——把那段报错发个 issue,就能把这个国家的流程补上
准备
Node 20+,以及 OpenCLI(在 1.8.6 上测过):
npm i -g @jackwener/opencli按 OpenCLI 的说明装好 Chrome 的 Browser Bridge 扩展,确认
opencli doctor通过。在这个 Chrome 里登录 ubereats.com,设好配送地址。
拿到本仓库,把 adapter 同步到 OpenCLI(OpenCLI 不认软链目录,必须复制过去;以后改了 adapter 也要再跑一次):
git clone https://github.com/Saekisui/ubereats-mcp.git && cd ubereats-mcp npm run sync先用命令行试一下,都是只读的:
opencli ubereats whoami opencli ubereats search ramen --limit 5想用日文界面就在命令前加
UE_LOCALE=jp。想让 AI 能下单的话,编辑
policy.json,把"enabled"改成true,按自己的情况调上限(账号币种),需要的话写上"timezone"(如"America/New_York")。不改就只能搜、看、加购,不能下单。注册 MCP。以 Claude Code 为例:
claude mcp add ubereats -- node /绝对路径/ubereats-mcp/mcp/server.js要日文界面就加
-e UE_LOCALE=jp;如果 opencli 不在~/.npm-global/bin/opencli,再加-e OPENCLI_BIN=/path/to/opencli。
一单怎么走
你:「帮我点份拉面」
AI:ue_search → ue_store(看菜单)→ ue_add(选好规格)→ ue_cart
AI:ue_review → 把快照原样给你看(店 / 品 / 费用 / 合计 / 地址 / 支付)
你:「好」
AI:ue_place → ue_orders(确认下成、之后跟配送)工具 | 读 / 写 |
| 读:开关、上限、今天已花多少、登录状态 |
| 读 |
| 读( |
| 写:加购物车 |
| 读:每家店一辆车,给出 draft_id |
| 读:结算快照 + confirm_token |
| 写·花钱 |
| 读 |
每次调用都会记一行到本地的 ledger.jsonl(已 gitignore),当日上限按它来算。
不经过 MCP,直接用命令行
opencli ubereats store <店id> --query ポカリ
opencli ubereats add <商品id> --spec "コーラ ポテト(L)" --qty 1
opencli ubereats cart
opencli ubereats review --draft <draft_id>
opencli ubereats place --draft <draft_id> --dry-run # 走到结算页、做完校验就停,不点下单
opencli ubereats orders已知限制
日本 + 日文界面实测最多;英文界面的下单步骤、其他国家的结算流程(比如美国的小费)还没人真跑过,认不出就会停下。Uber 的页面和接口随时会变,坏了先看 PITFALLS.md。
请求别太密。 Uber 有 reCAPTCHA 风控:请求太频繁时,接口会回 403(
botdefense: challenge)。这时去 Chrome 里打开 ubereats.com 手动搜一下、过了验证再用;脚本不会、也不该去碰验证码。浏览器命令请串行跑,别并发。第一次真下单,请人在旁边看着。
药品之类可能弹问诊或确认框,脚本认不出就会停下,需要你自己在 app 里完成。
License
AGPL-3.0,附加一条署名条款(见 NOTICE):包含或基于本项目的作品,必须同样以 AGPL-3.0 开源(做成网络服务给别人用也算),并保留 NOTICE 里的署名。
Available Tools
8 toolsue_addA
Add one item to the cart (a write action). spec = option keywords, space-separated, one per option group (e.g. 'Coke Large' / 'コーラ ポテト(L)'; multi-level options are matched level by level); qty = quantity 1-10. Confirm with ue_cart afterwards.
Notes:
Uber Eats keeps one cart per store; several carts can exist at once.
Adding is not buying. Confirm with the user before adding anything.
Inputs: id (required), qty, spec.
ue_cartA
List the carts (one per store): draft_id, store, items, most recently changed first.
Use this when:
You've just added something and want to confirm it landed, or need the draft_id for ue_review.
Inputs: no named parameters.
ue_ordersA
Active orders and delivery progress.
Use this when:
Right after ue_place, to confirm the order landed — always when ue_place came back unknown or with an error.
While the food is on the way, if the user asks where it is.
Inputs: no named parameters.
ue_placeA
Place the order. Requires the confirm_token from ue_review; the checkout snapshot is re-read and must still match (total / store / cart) and be within the per-order / per-day caps — otherwise it refuses.
Use this when:
The user has explicitly said yes to the exact review you showed them. Only then.
Notes:
If it comes back status unknown, or with any error after the gates passed, that does NOT mean it failed — the order may well have gone through. Check ue_orders first and tell the user what you find.
Never call ue_place again for the same cart: a second press is a second order and a second charge.
Inputs: confirm_token (required).
ue_policyA
Read the ordering policy: whether ordering is enabled, per-order / per-day caps (in the account's currency), how much was already spent today, and whether the user is logged in to Uber Eats in Chrome.
Use this when:
You're about to add to cart or place an order — check the caps first.
Inputs: no named parameters.
ue_reviewA
Checkout snapshot for one cart — the only trusted price: store / items / options / subtotal / every fee / total / ETA / delivery address / payment method. Read from Uber's checkout API keyed by draft (not the page), with two independent anchors that throw instead of mixing two carts. Returns a confirm_token (valid ~10 min, single use).
Use this when:
Before placing an order: show the user this exactly as returned; only after they say yes, call ue_place.
Inputs: draft.
ue_searchA
Search Uber Eats (delivers to the account's current address; works in any country Uber Eats serves). Works with store names, dishes and products. Product searches (e.g. 'Gatorade', 'ポカリスエット') return supermarkets / drugstores together with the matching items and prices.
Use this when:
The user asks for delivery ('点个 X' / 'order me X'). Delivery is real food and real money — don't start a delivery order on your own initiative.
Inputs: query (required), limit.
ue_storeA
A store's menu: item ids, names, prices, sections, quick-add, sold out.
Notes:
Supermarkets / convenience stores / drugstores only show a few shelves on their front page — pass query to search inside the store. Their item ids are long (store:item:section:subsection); pass them to ue_add unchanged.
Inputs: id (required), query.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 8 tools
Each tool maps to a distinct stage of the ordering workflow (policy, search, store, add, cart, review, place, orders) with no overlapping purposes. The descriptions reinforce the intended sequence, so an agent can reliably pick the right tool.
All tools share a consistent 'ue_' prefix and single-token names, which is highly readable. It mixes nouns (ue_policy, ue_cart, ue_orders) and verbs (ue_search, ue_add, ue_place) rather than a strict verb_noun pattern, but the convention is still predictable.
Eight tools map cleanly onto the browse → add → review → place → track lifecycle, with each tool earning its place. Nothing feels redundant or missing for the stated ordering purpose.
The core ordering lifecycle is fully covered with safety gates (policy, review, place, orders). Minor gaps exist: there is no way to remove/modify items from a cart, clear a cart, or cancel an order, which an agent cannot fully work around.
Maintenance
Related MCP Connectors
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Let AI agents work in your own signed-in Chrome, with approvals and verified results.
DoorDash MCP server for ChatGPT, Claude and any AI agent: search stores, read store menu, add item to cart, preview order total with fees, submit order, list past orders, get order status and reorder a past order in your own DoorDash account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to DoorDash yourself; Cotlet does not store your password and you need no DoorDash API key. About the official DoorDash API: DoorDash has no public API for ordering as a consumer (checked 27 Oct 2026). Its public APIs, Drive and Marketplace, are for merchants and delivery partners. Setup and FAQ: https://cotlet.dev/integrations/doordash-api-mcp-server
Pickup orders at Japanese restaurants for AI agents: find stores, read menus, place orders. No auth.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to securely control a user's existing signed-in Chrome browser through isolated tab groups, with strict per-session ownership and no cookie or token exposure.6 npmISC
- AlicenseBqualityBmaintenanceGives agents an autonomous browser that takes a single goal-level call and drives a real Chrome step by step, returning a verified result or a concrete question instead of guessing. It keeps passwords out of prompts, requires approval before buy/pay/delete actions, and detects login walls and bot challenges.920 npm2MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to drive the Chrome, Brave, or Edge browser you are already logged into, working in a tab of its own with a visible cursor and real trusted input. You can pause, take over, or hand back control at any time, and consequential actions such as submitting, paying, sending, deleting, or uploading wait for your approval.1Apache 2.0

K-Browserofficial
AlicenseNot gradedqualityBmaintenanceEnables local AI agents on macOS to drive the user's own signed-in Chrome through their own isolated tabs, with per-client tokens, blocked site categories, masked secrets and human hand-offs. Agents can open and navigate tabs, read pages, send real clicks and typing, inspect network requests, and hand tabs back to the user, all without touching other tabs or stealing focus.290 npm1MIT