Skip to main content
Glama

Munim on Alexa+

Ask your restaurant's books out loud and get answers you can trust. Munim is an operating system for small restaurant groups, bakeries and cloud kitchens: till, kitchen, stock, purchasing and ledgers, with AI specialists that explain what happened, find money leaking, and draft actions. A person approves every change.

This repo puts Munim behind Alexa+ as a self-hosted MCP server (Streamable HTTP, spec 2025-11-25) plus an Agent Skill, with a simulated Echo Show client. Alexa+ can already reorder a meal from a delivery app. What it can't do is answer "which ingredients are running low at Hazratganj?" from your own stock records, with every number cited, or say "Cannot verify" instead of guessing. It also can't take the owner's yes or no on a proposal. This repo adds those.

Tools

Tool

Who

What

shop_brief

owner

open findings (waste spikes, lapsing regulars, licences), how many are high, approvals waiting

ask_munim

owner

any question about the business. Munim's specialists answer only from the books: every number is a ⟨snapshot·row·field⟩ ref resolved by Munim core, returned as citations, or the answer abstains

pending_approvals

owner

the proposals Munim drafted (campaigns, purchase orders, recipe changes), numbered

approve_draft / reject_draft

owner

the owner's decision, recorded through Munim's guarded approval command (drafts.approve.v1 / drafts.reject.v1). The MCP layer adds no write path of its own

item_availability, menu, pickup_slots

customer

from the live storefront (public storefront key)

Every result carries a ready-to-speak speech sentence plus structured data for Echo Show cards.

Related MCP server: agent-storefront

How it works

Echo Show sim (browser MCP client)  ─┐
Claude / any MCP client             ─┼─ JSON-RPC over Streamable HTTP ─▶  api/mcp.js (Vercel function, stateless)
                                     ┘                                        │  lib/munim.js
                                                                              ├─ owner lane: Cobnut demo sign-in → Bearer + X-Munim-Tenant
                                                                              └─ customer lane: X-Munim-Storefront-Key
                                                                                   ▼
                                                              Munim core API (munim-api.vercel.app) → Postgres, AI specialists
  • api/mcp.js: McpServer + StreamableHTTPServerTransport (TS SDK 1.32.1, LATEST_PROTOCOL_VERSION 2025-11-25), stateless with JSON responses, one server per request.

  • lib/munim.js: the API client, citation resolver (_minor → ₹, _milli → units, _bp → %) and UUIDv7 command ids.

  • public/index.html: the simulated Echo Show. It's a real MCP client: initialize → tools/list → tools/call. It takes Web Speech voice in/out, renders cards and shows the wire log. Owner and customer modes.

  • skills/munim-owner/SKILL.md: the Agent Skill (intent → tool, citation and approval rules).

Run

npm install
MUNIM_STOREFRONT_KEY=… node scripts/local.mjs            # http://localhost:3000 (page) + /mcp
node scripts/check.mjs https://munim-alexa.vercel.app/mcp   # asserts protocolVersion 2025-11-25, lists tools, calls shop_brief
claude mcp add --transport http munim https://munim-alexa.vercel.app/mcp   # or MCP Inspector

Env: MUNIM_STOREFRONT_KEY (required for customer tools), MUNIM_API_ORIGIN, MUNIM_WEB_ORIGIN (defaults: the live Munim deployment).

Built during the hackathon

Munim itself (core API, AI specialists, storefront, Cobnut demo) existed before the hackathon. This repo (the MCP server, the citation resolver, the voice-first tool design, the Echo Show client and the Agent Skill) was built on 23 Oct 2026.

MIT licence.

AWS adapter

adapters/aws/ holds AWS counterparts of Munim's cloud adapters, four in all: idp-cognito (Amazon Cognito user pools in place of Identity Platform: ID-token verification against the pool's JWKS, plus the InitiateAuth password grant), blob-s3 (Amazon S3 in place of Cloud Storage for the blob port), llm-bedrock (Amazon Bedrock Converse API in place of Gemini / Vertex AI model calls, with per-role model ids) and db-rds (Amazon RDS / Aurora PostgreSQL in place of Supabase / Cloud SQL Postgres: a verified-TLS pg config with optional IAM auth tokens). Tests run offline: node --test 'adapters/aws/**/*.test.js'. See adapters/aws/README.md.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      A
      quality
      D
      maintenance
      An MCP server for restaurant discovery and booking across Resy and OpenTable via natural language. It integrates Google Places data with dietary preferences, visit history, and weather awareness to provide personalized dining recommendations and group reservation management.
      23
      -
    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables any AI agent to discover, query, and order from a restaurant's storefront via MCP tools. It handles menu lookup, modifier validation, and enforces a mandatory confirmation gate before payment, replacing the human-operated phone line.
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to search restaurants, publish and maintain store data, submit verified freshness feedback, and query demand or updates through MCP tools. It provides both read-only and full administrative endpoints backed by open restaurant data.
      1
      AGPL 3.0