Moodish: say the mood, get the meal

Moodish concierge

What would hit the spot?

Dish, texture, weather, energy. Say as much or as little as you know.

Hey, I’m Moodish.

Tell me what you feel like eating. Try “chewy smoky chaap” or give me everything at once: “spicy Chinese, non-veg, under ₹450.”

I’m checking the time of day so I can make the first question useful.

How adventurous?

Maximum spend₹450 ₹150₹1,000

Choose once, or include these details in your message.

Moodish Enterprise

Lunch operations, without the spreadsheet.

A manager sets the policy. Employees respond privately. Moodish builds one efficient, approval-ready plan.

Enterprise controls · private by default

Manager sets policyHeadcount, budget per person, delivery need, and decision method.

Employees answer privatelyA web link or workplace app collects cravings, diet rules, and allergies.

Moodish consolidatesOne restaurant first; a split is disclosed only when hard constraints require it.

Creator confirms the cartThe manager approves a plan; the Swiggy-connected creator confirms the final mutation.

Invite employees privately

Launch a lunch to create the private response link.

Workplace app adapters

Deploy where employees already work.

Each adapter starts the same private enterprise workflow. Public channels see progress and aggregate coverage, never individual dietary or allergy details.

Moodish for Slack/moodish starts the lunch; teammates open a private form.Adapter ready · app credentials required

Moodish for Microsoft TeamsAn adaptive card launches collection and reports safe progress.Adapter ready · app registration required

Moodish for DiscordA command starts collection; private responses feed the same dashboard.Adapter ready · bot credentials required

How do I test Moodish in Slack?

Create an app at api.slack.com/apps and choose “From scratch.”
Add a slash command named /moodish. Set its Request URL to https://moodish.onrender.com/api/platforms/slack/events.
Copy the app’s Signing Secret to Render as SLACK_SIGNING_SECRET.
For the manager dashboard, add the redirect URL https://moodish.onrender.com/api/platforms/slack/oauth/callback, then set SLACK_CLIENT_ID and SLACK_CLIENT_SECRET on Render.
Install the app to your workspace, redeploy Moodish, then type /moodish easy team lunch in a test channel.

The channel receives safe progress plus a private preference link. The manager button uses Slack identity to protect approval access.

Product + engineering reference

From a mood to a guarded cart.

This page documents what Moodish does today, which calls are conditional, and which capabilities are deliberately not implemented.

AI semantic intentDeterministic rankingFood cart needs confirmationCheckout disabled

System boundary

The web experience, workplace adapters and Moodish MCP share one agent service. Fixture and live data are normalized before the same ranking engine.

Moodish system overview diagram — Download editable Excalidraw source

Product surfaces

Solo chat, enterprise dashboard, private participant link, Slack/Teams/Discord adapters and Moodish MCP tools.

Agent core

Conversation state, intent extraction, hard constraints, ranking, recommendation trace and confirmation gates.

Persistence

PostgreSQL when DATABASE_URL is configured; process-memory fallback for local fixture runs. Stores profiles, confirmed plans and group sessions.

User journeys

Solo asks only for missing hard fields. Enterprise gathers sensitive constraints privately and exposes aggregate coverage to the manager.

Solo and enterprise Moodish user journey diagram — Download editable Excalidraw source

solomood → clarify diet/budget → discover → rank → select → optional Instamart preview → explicit Food cart confirmation enterprisepolicy → private preferences → lock → rank → manager/vote/automatic → creator confirms Food cart

AI boundary

The language model extracts structured craving semantics, recognises edits to an active plan, and writes friendly copy. It does not select Swiggy tools, apply dietary filters, mutate carts or set ranking scores.

Diagram showing AI semantic interpretation and the deterministic Moodish safety boundary — Download editable Excalidraw source

Output guard

AI intent is schema-constrained and then validated; a deterministic extractor remains available during provider failure. If the AI summary contradicts ranking, Moodish replaces it. Current request, dietary rules and availability always outrank model phrasing, profile history or discovery preference.

Swiggy OAuth + MCP calls

These names and argument shapes are taken from the current gateway implementation. Every applicable discovery and cart call carries the selected addressId.

Detailed Swiggy OAuth and MCP call lifecycle diagram — Download editable Excalidraw source

When	Server / tool	Arguments sent	Effect
First Swiggy connect	`POST /auth/register`	redirect URIs, client metadata	Dynamic Client Registration
OTP consent	`GET /auth/authorize`	`client_id`, redirect, state, PKCE challenge	User authorization
Callback	`POST /auth/token`	authorization code, verifier, redirect	Encrypted access-token session
Start planning	`food · get_addresses`	`{}`	Read saved delivery addresses
Explicit dish	`food · search_menu`	`{ addressId, query, vegFilter }`	Exact dish search; may run again for a labelled alternative
Vague mood / cuisine	`food · search_restaurants`	`{ addressId, query }`	Restaurant discovery; a broad empty query is a last candidate-pool fallback
Candidate hydration	`food · get_restaurant_menu`	`{ restaurantId, addressId }`	Read menu and current availability
Optional pairings	`im · search_products`	`{ addressId, query }`	Read Instamart products; no cart write
After explicit approval	`food · update_food_cart`	`{ restaurantId, addressId, items: [{ itemId, quantity }] }`	Mutates Food cart only

POST https://mcp.swiggy.com/food
Authorization: Bearer <encrypted-session-token>

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "search_menu",
    "arguments": { "addressId": "...", "query": "chaap", "vegFilter": 1 }
  }
}

Transport retries rate limits, 5xx responses and transient upstream failures up to four attempts. Expired or invalid OAuth sessions require reconnection.

Ranking, pairings and memory

Recommendation quality is ordered by hard constraints first. Profile history is deliberately weak context.

Hard filtersavailability, veg/non-veg, dietary rules and allergies
Cravingexact dish, texture, flavour, preparation and cuisine relevance
Commercial fitmaximum budget as a ceiling, then distance/delivery suitability and rating
Tie-breakcomfort, balanced or explore; never allowed to override a dish match

Pairings

Context rules create search queries such as Chinese → chilled drink, biryani → raita and spicy food → cooling drink. Results must be available and fit remaining budget.

Meal memory

Stores confirmed Food items plus selected Instamart preview items and totals. It is not proof that a Swiggy order was placed or delivered.

Separation

Food and Instamart remain different fulfilment flows. Today only the Food cart has a mutation tool in Moodish.

Deliberately not claimed

A capability is not presented as live until its schema and authenticated staging journey pass.

Swiggy checkout or order placement
Swiggy completed order-history ingestion
Instamart cart mutation
Full pagination traversal
Variant and restaurant add-on resolution
Payment selection or delivery scheduling

Tell Moodish what works for you.

What would hit the spot?

Here’s what fits.

Lunch operations, without the spreadsheet.