Peckish 🍜

Feeling peckish? Just ask. An AI ordering agent for DoorDash — it searches, compares real totals (fees included), builds the cart, and you approve every order. Built on Claude and DoorDash's official dd-cli.

One tool layer, three surfaces:

Which one is for me?

• Comfortable with a terminal? → Terminal chat. Fastest, most informative.
• Want something that feels like an app? → Local web app. Cards, live quote, a proper Place-order button.
• Already use Claude Desktop or Claude Code? → MCP server. No API key, no separate chat window — Claude itself becomes your ordering agent, and order confirmation appears as a native dialog.

Peckish finding a dinner under $25: tool calls run, a recommendation streams in with the fee-included total, and an order-confirmation dialog asks for approval before anything is placed.
_{Scripted demo of the web app — store, prices and card are fictional; the interface and the approval gate are the real ones.}

peckish on the web →

Install in one line — no git clone:

npm install -g peckish

Prefer an app? Download the Mac app (.dmg) — guided setup, no terminal at any step. See Mac app below. Everything runs on your own Mac either way, because DoorDash sign-in lives in your keychain.

you › Find me a high-protein dinner under $25 that can arrive within 45
      minutes. Avoid mushrooms and excessive fees.

⚙ search_restaurants {"query":"grilled chicken bowls","limit":8}  ✓ 2.1s
⚙ get_menu {"store_id":"35406455","filter":"chicken"}             ✓ 1.8s
⚙ list_carts {"store_id":"35406455"}                              ✓ 1.2s
⚙ add_items_to_cart {…}                                           ✓ 2.4s
⚙ preview_order {"cart_uuid":"…"}                                 ✓ 3.9s

Best fit: Sharon Korean Kitchen (4.8★, ~24 min) — Grilled Chicken Bulgogi
Bowl, $16.95. No mushrooms listed. Total with fees: $21.40 on your Visa
ending 1234. Suggested Dasher tip is $3.50 — that, another amount, or none?
~$0.04 turn · $0.04 session

Get started (user guide)

1. Prerequisites

• A Mac with Apple Silicon (M1–M4). Peckish is local-first: your Mac is the backend on every surface, because dd-cli authenticates against your keychain.
• Node.js 20+ — node --version to check; install from nodejs.org or brew.
• DoorDash CLI access (currently waitlist-gated by DoorDash). Download the release from doordash-oss/doordash-cli, verify the SHA256 checksum against the published value, then:
  sh

  tar -xzf dd-cli-v*-darwin-arm64.tar.gz && cd dd-cli-v*-darwin-arm64
  bash install.sh          # installs to ~/.local/bin/dd-cli
  dd-cli login             # sign in to DoorDash in your browser (stored in keychain)
  

• An Anthropic API key for the terminal/web surfaces (console.anthropic.com) — or skip the key entirely and use the MCP surface with your Claude subscription (step 4).

2. Install Peckish

npm install -g peckish

That's it — you now have three commands: peckish (terminal chat), peckish-web (web app), and peckish-mcp (MCP server).

git clone https://github.com/CydVilla/peckish.git
cd peckish
npm install
npm test          # optional: 13 unit tests, no network needed
npm run dev       # terminal chat (or: npm run web / npm run mcp)

3. Run it — terminal or web

export ANTHROPIC_API_KEY=sk-ant-…   # from console.anthropic.com

peckish         # terminal chat
peckish-web     # web app → open http://localhost:4747

On boot Peckish verifies your DoorDash sign-in, shows your default delivery address, and flags any open carts you forgot about. If sign-in expired, run dd-cli login again and restart.

4. Or run it inside Claude — no API key

Claude itself becomes the ordering agent, and your Claude subscription pays for the model. Pick whichever fits your client:

Claude Code — one line:

claude mcp add peckish -- npx -y peckish-mcp

Claude Desktop — download peckish-0.2.2.mcpb from Releases and double-click it. Claude Desktop installs it like a browser extension: no terminal, no Node install, no JSON editing.

{
  "mcpServers": {
    "peckish": {
      "command": "npx",
      "args": ["-y", "peckish-mcp"]
    }
  }
}

Peckish is also listed in the MCP Registry as io.github.CydVilla/peckish, so clients that browse the registry can find it directly.

Restart Claude Desktop and ask it to find you dinner. Order confirmation appears as a native dialog; clients that can't render dialogs can browse and build carts but cannot place orders (fail closed).

5. Everyday use

Things to say:

• "Find me a high-protein dinner under $25 that can arrive within 45 minutes. Avoid mushrooms and excessive fees."
• "Compare the real totals at the top two — fees included."
• "Reorder my usual from Sharon Korean."
• "What did I spend on delivery last month?"
• "Is this place actually good?" (checks web reviews)
• "Never mushrooms, ever." → saved permanently; applied automatically next time
• "Get me milk, eggs, and a pound of ground beef from Whole Foods."

Placing an order always ends with an explicit confirmation you perform — typing yes in the terminal, clicking Place order in the web modal, or approving the dialog in Claude Desktop. Before that, Peckish must show you the itemized quote, confirm the tip, and name the card being charged. Decline anything and it backs off.

Controls & housekeeping

Cost: defaults are tuned for low spend at decent quality — claude-sonnet-5 at medium effort, prompt caching on the system prefix and conversation tail, and server-side context editing that prunes stale menu payloads in long sessions. The cost meter shows the approximate spend per turn and per session. Max quality: DD_AGENT_MODEL=claude-opus-4-8 DD_AGENT_EFFORT=high. (On MCP, the client chooses and pays for the model.)

Troubleshooting

Env vars: DD_AGENT_MODEL (default claude-sonnet-5), DD_AGENT_EFFORT (low–max, default medium), DD_CLI_PATH, PECKISH_PORT (default 4747).

Mac app

A double-clickable app for people who never want to see a terminal: download Peckish-x.y.z-arm64.dmg from Releases, drag Peckish to Applications, and open it.

First launch (Gatekeeper): the app is ad-hoc signed but not notarized (no paid Apple Developer ID), so macOS won't open it on a plain double-click the first time. Right-click the app → "Open" → "Open" (or approve it under System Settings → Privacy & Security). Only needed once.

If macOS instead says "Peckish is damaged and can't be opened", you have a build from before this was fixed, or the download quarantine got confused. Clear it once and it opens normally:

sh

xattr -cr /Applications/Peckish.app

First-run setup happens in the app — three buttons, no terminal:

1. DoorDash CLI — one-click guided install (downloads the official release, verifies its SHA256 checksum before running anything). If you don't have dd-cli access yet, there's a waitlist link.
2. Sign in to DoorDash — opens your browser; the app detects when you're done. Your sign-in lives in the macOS keychain.
3. Anthropic API key — paste it once; it's stored encrypted with Electron safeStorage (keychain-backed), never in plain text.

Then Open Peckish — same web app, same order-confirmation modal, same safety gates; the app is just a shell that runs the local server for you on a random localhost-only port. Requires Apple Silicon; Node.js is not required (the app bundles its own runtime).

Building it yourself: cd desktop && npm install && npm run dist → desktop/dist/Peckish-*.dmg.

What it does

• Search → menus → cart → preview → confirm → submit, with the real fee/ETA quote (order preview) driving every recommendation.
• Comparison shopping: builds carts at up to 3 finalists, compares true totals + fee share + ETA, recommends one, deletes the losers.
• Fee tactics: promo scanning with consent, pickup-vs-delivery comparison, DoorDash credits surfaced, DashPass status shown.
• Memory: dietary rules and habits persist across sessions and surfaces.
• History: "my usual" from order frequency, honest spend breakdowns from receipts, reorders with silent-drop detection.
• Web reviews via Claude's server-side web search (never used for prices — dd-cli is the only source of truth for ordering data).
• Work benefits (company budgets + expense codes), scheduled delivery, pickup, groceries/retail/pets/alcohol.

Architecture

terminal REPL          local web app           MCP client (Claude Desktop…)
 src/index.ts           src/web.ts + public/    src/mcp.ts
      └────────────┬─────────┘                       │  (client's model reasons;
                   ↓                                 │   server instructions guide it)
     Claude agent loop, streaming                    │
     src/agent.ts · claude-sonnet-5 · strict tools   │
     · adaptive thinking · context editing           │
     · prompt caching · web_search · cost meter      │
                   └───────────────┬─────────────────┘
                                   ↓
                 28 typed tools — src/tools.ts (strict: true)
                                   ↓
                 sanitizing wrapper — src/ddcli.ts
                                   ↓
                 dd-cli --json-output  →  DoorDash

Safety model

Placing an order always requires an explicit human approval rendered by the surface, not by the model:

Also on every surface:

• Strict tool schemas — the API guarantees tool arguments validate before any handler runs (no malformed-argument class).
• Abortable turns — Ctrl+C / Stop rolls history back to the turn start.
• Audit log — every tool call, argument set, duration, confirmation outcome, and submit result in ~/.peckish/logs/*.jsonl.
• Tip confirmed + card named before any submit ask; submit never auto-retries (not idempotent); success reported only after order status confirms.
• Merchant text treated as data (widget/assistant-instruction fields stripped); read-only CLI calls retry once on transient errors, mutations never do.
• Web server is localhost-only (Host + Origin checks).

Repo map

Releasing

Everything ships from one tag. .github/workflows/release.yml publishes both npm packages, registers the MCP Registry entry, builds the .mcpb and the .dmg, and attaches both to the GitHub release:

npm version patch          # or edit the versions by hand
git push && git push --tags

Every publish step is skip-if-already-published, so re-running a tag after a failure is safe. .github/workflows/ci.yml runs typecheck, tests, a metadata consistency check, and a real MCP handshake on every push.

scripts/check-consistency.mjs guards the metadata that spans files and drifts silently — the registry namespace casing, matching server.json name and mcpName, the 100-character registry description cap, and the extension's advertised tool list. Run it locally before tagging.

Notes & limitations

• Local-first by design: hosted delivery (SMS bots, voice) would require DoorDash's partner API — your Mac is the backend here.
• One open cart per store (DoorDash rule) — Peckish collision-checks and asks.
• payment-method list sees cards only; wallet defaults (Apple Pay etc.) are confirmed generically or via the browser checkout URL.
• Age-restricted items can't be submitted by an agent — checkout URL fallback.
• Popularity data is deliberately unused (per dd-cli guidance); web reviews fill that gap with attribution.
• Cost figures are close estimates from token usage at list prices.