Udhaar

A voice-first credit ledger (udhaar/khata) for small shopkeepers, built as a self-hosted Alexa+ MCP server over Streamable HTTP, backed by Amazon Bedrock for understanding mixed-language speech and drafting collection messages.

Built for Amazon's Build, Ship, Shape hackathon — Alexa+ track, plus the AWS Builder and Open Source mini-challenges.

The problem

Millions of small Indian shops run on a paper notebook (khata) of who owes what. It's error-prone, hard to search, impossible to back up, and useless hands-free when a shopkeeper is mid-sale with both hands full of goods. A voice assistant that can record "Ramu took two hundred rupees of rice on credit" while the shopkeeper keeps working — and later tell them who to follow up with — replaces that notebook without asking anyone to learn an app.

Related MCP server: FarmAssist Agent

The demo in 30 seconds

Open the web simulator (GET /, see Setup) and speak or type:

1. "Ramu took two hundred rupees of rice on credit and Kumar paid five hundred" → both entries recorded in one sentence, parsed by Bedrock (or a regex fallback if Bedrock isn't configured).

2. "How much does Ramu owe?" → spoken balance.

3. "Who should I follow up with today?" → ranked list with a drafted, polite reminder message per customer.

4. The ledger panel on the right updates live after every action.

Why this isn't a thin wrapper

• Bedrock-backed understanding with a graceful fallback. Spoken, code-mixed sentences (English + transliterated Tamil/Hindi — udhaar, baaki, kadan) are parsed by a Bedrock ConverseCommand call (src/parser.ts). If Bedrock is unavailable, misconfigured, or disabled (UDHAAR_DISABLE_BEDROCK=1), a regex-based rules engine takes over automatically — the tool never crashes or goes silent, it just reports which engine it used (structuredContent.engine).

• Proactive collection planning, not just a balance lookup. collection_plan ranks customers by balance × recency and drafts a warm, non-threatening reminder per customer — in the shopkeeper's requested language — via Bedrock, again falling back to a plain template if Bedrock is unavailable.

• Multilingual by design, not just English with a translation layer bolted on: the Bedrock prompts are written to understand code-mixed Tamil/Hindi/English input directly, and collection_plan can draft output in a requested language.

• Voice-shaped responses. Every tool returns one or two short spoken sentences (plus structuredContent for UIs) — not JSON dumps, not markdown tables, because the Alexa+ assistant on the other end has to read it aloud.

• A real, spec-compliant MCP server, not a demo stub: stateless Streamable HTTP per MCP 2025-11-25, validated against the official @modelcontextprotocol/inspector CLI and the official SDK client end-to-end (src/smoke-test.ts), with bearer-token auth and a Docker/App Runner deployment path.

Architecture

flowchart LR
    subgraph Client
        Alexa["Alexa+ (voice)<br/>— unverified, see docs/ALEXA_SETUP.md"]
        Sim["Web simulator<br/>(public/, this repo)"]
        Inspector["MCP Inspector / SDK client"]
    end

    subgraph Server["Express server (src/server.ts)"]
        MCP["POST /mcp — Streamable HTTP<br/>(stateless, MCP 2025-11-25)"]
        Ask["POST /api/ask — intent router<br/>(src/intent.ts)"]
        State["GET /api/state"]
        Health["GET /health"]
    end

    subgraph Logic
        Parser["src/parser.ts<br/>parseUtterance()"]
        Planner["src/planner.ts<br/>collectionPlan()"]
        Ledger["src/ledger.ts<br/>JSON-file store"]
    end

    Bedrock["Amazon Bedrock<br/>ConverseCommand"]

    Alexa -->|Bearer auth| MCP
    Inspector -->|Bearer auth| MCP
    Sim --> Ask
    Ask -->|self-call, same auth| MCP
    MCP --> Parser
    MCP --> Planner
    MCP --> Ledger
    Parser -.fallback to regex.-> Parser
    Parser --> Bedrock
    Planner --> Bedrock
    Planner --> Ledger

Tools

Setup and run

npm install
cp .env.example .env    # fill in what you need; see below
npm run dev              # tsx src/server.ts, serves on :3000

Open http://localhost:3000/ for the web simulator, or point an MCP client at http://localhost:3000/mcp.

Everything works with no AWS credentials at all if you set UDHAAR_DISABLE_BEDROCK=1 — speech parsing and reminder drafting fall back to deterministic regex/template logic, useful for offline development or a no-AWS demo.

Bedrock setup

1. In the Bedrock console for your target region, go to Model access and request access to an Anthropic Claude model (usually instant approval).

2. Find a valid, invokable model ID for your region/credentials:

   aws bedrock list-inference-profiles --region <your-region> \
     --query "inferenceProfileSummaries[].inferenceProfileId"

   Many current Claude models require an inference profile ID (us.anthropic.claude-..., eu.anthropic.claude-..., global.anthropic.claude-...), not the raw foundation-model ID.

3. Set in .env:

   AWS_REGION=<your-region>
   BEDROCK_MODEL_ID=<the inference profile ID from step 2>

4. Standard AWS credential chain only (aws configure, AWS_PROFILE, instance/task role) — no credentials are ever hardcoded.

Environment variables

See .env.example for the full, documented list (PORT, AWS_REGION, BEDROCK_MODEL_ID, UDHAAR_API_KEY, UDHAAR_DISABLE_BEDROCK, UDHAAR_DATA_FILE).

Tests

npx tsc --noEmit    # type check
npm run test:unit   # node:test unit tests — ledger, parser, planner, intent router
npm test             # end-to-end: spins up a real server, drives it with the
                      # official MCP SDK client over Streamable HTTP, asserts
                      # on actual tool results (not just "it responded")

Deploy

Docker + AWS App Runner, with an IAM instance role scoped to exactly bedrock:InvokeModel, a health-check wired to /health, and a cost/shutdown note: see deploy/README.md. The Docker image has been built and smoke-tested locally (non-root user, /health and /mcp both verified inside the container); the App Runner deployment steps themselves require an AWS account and have not been executed by anyone but you.

Security notes

• No hardcoded credentials anywhere; AWS access only via the standard SDK credential chain.

• /mcp and /api/* require Authorization: Bearer <UDHAAR_API_KEY> when that variable is set (constant-time comparison); unset means auth is open — set it before exposing the server publicly. /health always stays open for load balancer health checks.

• The bearer-token scheme here is not full OAuth 2.1 — Alexa+'s documented requirement for real device registration is OAuth 2.1 with PKCE, which this repo does not implement (see docs/ALEXA_SETUP.md for the honest gap analysis).

• The JSON-file ledger store is demo-grade (not concurrent-write-safe, not durable across container restarts on App Runner) — see the DynamoDB note in deploy/README.md for the production path.

What's verified vs simulated

• Verified: Streamable HTTP MCP server behavior (Inspector CLI + the official SDK client, both over real HTTP), bearer auth (401/200 paths), Bedrock parsing/drafting with regex/template fallback, the Docker image running as non-root locally.

• Simulated: the "Alexa+" experience is demonstrated via the included web simulator (public/), which speaks to the same /mcp tools over the same Streamable HTTP protocol an Alexa+ registration would use — browser speech recognition/synthesis stand in for the device. Real Alexa+ device registration has not been completed; see docs/ALEXA_SETUP.md for exactly what that would take and why it's marked unverified.

Hackathon mapping

• Alexa+ track (primary): self-hosted MCP server, Streamable HTTP, MCP spec 2025-11-25, voice-shaped tool responses.

• AWS Builder mini-challenge: Amazon Bedrock (ConverseCommand) for speech parsing and message drafting; AWS App Runner + ECR deployment path.

• Open Source mini-challenge: see the project's hackathon submission notes for the external contribution made during the hackathon window (tracked separately from this repo's own commits).

License

MIT — see LICENSE.