Add Express + Strands TypeScript chatbot starter (SSE, Zod tools, MCP, interventions)

[hsaenzG]

Summary

Adds a new TypeScript sample: an Express + Strands Agents chatbot starter at typescript/01-learn/07-express-chatbot.

It rebuilds the shape of a typical chatbot service (Node.js, Express, three tools) on the Strands Agents TypeScript SDK, and demonstrates several SDK features together in one runnable service:

• SSE streaming with event ids — POST /chat/stream forwards every AgentStreamEvent as its own Server-Sent Event with an incrementing id: line (plus a leading session event and a trailing done/error event).
• Non-streaming route — POST /chat returns the final reply as JSON, plus GET /health.
• Zod tools — three custom tools (get_current_time, calculate, get_weather) with Zod input schemas.
• Session store — in-memory per-session conversation history, replayed into the agent each turn.
• MCP client — connects to an existing remote MCP server over Streamable HTTP and loads its tools alongside the local ones. Defaults to the public, no-auth AWS Knowledge MCP Server, so the agent can search and read official AWS documentation out of the box.
• Interventions — a rate-limit intervention (beforeToolCall) caps tool calls per session, and a moderation intervention (beforeModelCall) blocks configured terms.

The default model provider is Amazon Bedrock (Amazon Nova Pro). The sample follows the project structure and README template in .github/templates/use-cases/, includes an architecture diagram, and updates the 01-learn index with the new entry.

Project structure

typescript/01-learn/07-express-chatbot/
├── README.md
├── package.json
├── tsconfig.json
├── .env.example
├── images/
│   ├── architecture.png
│   └── architecture.svg
└── src/
    ├── index.ts                          # entry point: env, MCP connect, server start
    ├── server.ts                          # Express app: /health, /chat, /chat/stream (SSE)
    ├── agent.ts                           # per-turn agent factory
    ├── config.ts                          # env-driven configuration
    ├── tools/chatTools.ts                 # three Zod tools
    ├── sessions/SessionStore.ts           # in-memory session store
    ├── mcp/mcpTools.ts                    # MCP client (AWS Knowledge MCP by default)
    ├── interventions/RateLimitIntervention.ts
    ├── interventions/ModerationIntervention.ts
    └── smoke.ts                           # credential-free smoke test

Testing

Verified locally:

• npm run typecheck (tsc --noEmit) — passes with no errors.
• npm run smoke — passes: tool registration, moderation (proceed/deny), rate limiting (allow up to cap then deny), and the HTTP layer (400 on empty body, /health ok).
• Manual run: the server boots, connects to the AWS Knowledge MCP Server and loads its 5 tools, /health returns {"status":"ok"}, and POST /chat/stream emits well-formed SSE frames (id: / event: / data:) starting with the session event. The Bedrock model call requires valid AWS credentials and model access at runtime; failures surface as an SSE error event rather than crashing the process.

Notes

• Running the chatbot requires AWS credentials in the default credential chain and Bedrock access for the configured model (default us.amazon.nova-pro-v1:0).
• The MCP integration connects to the public AWS Knowledge MCP Server and needs no credentials; set MCP_SERVER_URL to another Streamable HTTP MCP server, or empty to disable.
• The moderation intervention uses simple substring matching for illustration and should be replaced with a real moderation service for production use (called out in the README disclaimer).

Tracking

This is a new feature, so there is no GitHub issue to close (no Fixes # applies). Tracked in Asana: https://app.asana.com/1/8442528107068/project/1217155663608611/task/1218296727005011