DoorDash MCP
Enables ordering DoorDash through Poke over text, exposing tools for restaurant discovery, cart management, promotions, and order submission with purchase safety features.
README
Poke × DoorDash CLI recipe
A Poke recipe that lets Poke order DoorDash for you over text,
powered by DoorDash's official dd-cli (limited beta, announced 2026-07-15 by
@andyfang).
Fork of DanteMagg/poke-doordash-recipe,
upgraded for dd-cli v0.2.1+ (mandatory --intent) and hardened around real
upstream failure modes (silently dropped cart items, no submit idempotency).
Poke installs recipe integrations as MCP servers, so this repo is a thin MCP
bridge: Poke → tunnel → server.mjs → dd-cli → DoorDash.
Prerequisites
- dd-cli — official binary from doordash-oss/doordash-cli
(macOS Apple Silicon / Linux x86_64). Download the latest release tarball,
verify the SHA256, and put the binary on PATH as
dd-cli(or setDD_CLI_BIN). Full functionality is waitlist-gated — join at https://forms.gle/gvCQZvu9C1EKA6aM6. - Node 22.5+ (the submission journal uses the built-in
node:sqlite). - A Poke account with Kitchen access.
- Log in once via
dd-cli login(interactive; credentials go to the OS keychain). Note: v0.2.2 gates even subcommand--helpbehind login.
Run + publish
npm install
npm test # 19 tests against a fake dd-cli; no account needed
npm start # MCP server on http://127.0.0.1:3737/mcp (loopback only)
npm run tunnel # in a second terminal: tunnels to Poke and creates the recipe
npm run tunnel runs npx poke@latest tunnel http://localhost:3737/mcp -n "DoorDash CLI" --recipe,
which prints a shareable recipe link + QR code. (Log in first with npx poke@latest login.
The tunnel only forwards the port — start npm start before it.)
If poke login is broken (cli-auth outage)
As of 2026-08-12, POST poke.com/api/v1/cli-auth/code returns an empty 200, so
poke login dies with "Unexpected end of JSON input" and web API keys don't
carry CLI scopes. Bypass the Poke CLI entirely:
openssl rand -hex 24 > .data/mcp-auth-token
MCP_AUTH_TOKEN=$(cat .data/mcp-auth-token) npm start # /mcp now requires the bearer token
cloudflared tunnel --url http://localhost:3737 # prints a public https URL
Then in Poke's web app (Settings → Connections → add an MCP integration): URL =
https://<cloudflared-host>/mcp, API key = the token. The quick-tunnel URL
changes each cloudflared restart — update the connection when it does.
Then polish in Kitchen → your recipe. Use RECIPE.md as the recipe's instruction text — it's the agent-facing operating manual.
Tools exposed
28 tools. Everything from the upstream bridge, plus search_products_across_stores:
- Discovery:
search_restaurants,find_nearby_stores(grocery/retail/alcohol/ convenience/pets),get_menu,find_items,search_products_across_stores(multi-store product hunt with deterministic diet/zero-sugar synonym expansion),restaurant_item_details,item_details,store_details,build_grocery_list - Account:
list_addresses,payment_methods,doordash_status(install / version / login / waitlist probe) - Cart:
cart_list(pre-flight before creating carts),cart_add_items,cart_show,cart_remove_item,cart_delete— all mutations are read-back verified (upstream can silently drop items) and returnverified: trueor acart_mismatchwith exactly what's missing - Promos:
promo_list,promo_apply,promo_remove - Orders:
order_history,order_reorder,order_preview(relays dd-cli's--beautifysummary verbatim and mints apreview_token),order_submit,order_status,order_receipt,checkout_url(browser fallback for payment-method swaps) - Escape hatch:
dd_cliraw passthrough (login,export-token, and raworder submitblocked)
Purchase safety
order_submitcharges the default card immediately. It requires a fresh single-usepreview_tokenfromorder_preview— the server re-reads the cart at submit time and refuses if the contents, fulfillment, or schedule changed since the preview — plususer_confirmed=true, legitimate only after the preview was shown, the total approved, the card NAMED (brand + last4 frompayment_methods), and the Dasher tip explicitly confirmed (in cents).- Duplicate submits are blocked durably (SQLite journal in
.data/state.db, override withDD_MCP_STATE_DB): a restart, a second server instance, or a Poke retry after a timeout cannot double-order. Ambiguous submits lock the cart and are never auto-retried. - Every dd-cli service call carries a mandatory
--intent(v0.2.1+) — always a deterministic description of the MCP tool being run, never the user's own words. CLIs without the flag are auto-detected. - Cart writes auto-spaced ~8s apart (DoorDash rate limiting).
- Errors map to a stable taxonomy (
not_logged_in,waitlist_pending,rate_limited,cart_mismatch,preview_expired,duplicate_submit_blocked,ambiguous_submit, …) with actionable messages the Poke agent can relay. - Server binds to loopback only by default (
HOSTenv to override); the Poke tunnel is the intentional external bridge. Structured logs carry tool/family/ duration/error-type only — no payloads, no tokens, no card data.
Environment variables
| Var | Default | Purpose |
|---|---|---|
DD_CLI_BIN |
dd-cli |
Path to the dd-cli binary |
PORT |
3737 |
MCP server port |
HOST |
127.0.0.1 |
Bind address |
DD_MCP_STATE_DB |
.data/state.db |
SQLite path for the submission journal |
MAX_RESPONSE_CHARS |
60000 |
Tool-response truncation cap |
DD_MCP_QUIET |
– | 1 silences structured logs |
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。