State Government Services Bot — Rasa Pro CALM + MCP Sub-Agents

A Rasa Pro chatbot demonstrating SNAP benefits eligibility checking and DFCS appointment scheduling. Built with the LLM-native CALM approach and MCP sub-agent architecture — no intents, stories, rules, or custom Python actions.

Capabilities

• SNAP eligibility check — Collects household size and income, calls an MCP tool via sub-agent, and returns eligibility results
• DFCS appointment scheduling — Sub-agent queries available slots, user confirms, and a direct MCP tool call books the appointment
• General SNAP info — Program overview with links to apply online
• Graceful boundaries — Out-of-scope requests (license, taxes, Medicaid, etc.) get a polite redirect

Quick Start

cp .env.example .env
# Add your RASA_PRO_LICENSE and OPENAI_API_KEY

uv sync
uv run rasa train

# Start the MCP server (separate terminal)
uv run mcp_server/server.py

# Launch the inspector
uv run rasa inspect

Architecture

User → Rasa (CompactLLMCommandGenerator + FlowPolicy) → Flow
  → snap_agent (sub-agent)            → mcp_server → check_snap_eligibility
  → appointment_selector (sub-agent)   → mcp_server → query_available_appointments
  → book_appointment (direct MCP call) → mcp_server → book_appointment

All service logic lives in the MCP server. Sub-agents are config-only (no custom Python). The bot uses Rasa's built-in response rephraser so templated responses sound natural.

Project Structure

├── config.yml                     # Pipeline: CompactLLMCommandGenerator + FlowPolicy
├── endpoints.yml                  # MCP server connection, rephraser config, model groups
├── domain/
│   ├── domain.yml                 # Slots and responses for SNAP + appointments
│   ├── chitchat.yml               # Greeting, goodbye, help responses
│   └── shared.yml                 # Shared slots and refusal responses
├── data/flows/
│   ├── snap_benefits.yml          # SNAP eligibility check + general info flows
│   ├── schedule_snap_appointment.yml  # Appointment query, confirm, and book flow
│   ├── unsupported.yml            # Catch-all for out-of-scope requests
│   └── chitchat.yml               # Greeting, goodbye, help, small talk flows
├── mcp_server/
│   └── server.py                  # FastMCP server with 3 tools (mock data)
├── sub_agents/
│   ├── snap_agent/config.yml      # MCPOpenAgent → check_snap_eligibility
│   └── appointment_selector/config.yml  # MCPOpenAgent → query_available_appointments
└── tests/                         # E2E test cases

Key Patterns

1. LLM-based routing — CompactLLMCommandGenerator reads flow descriptions and selects the right flow. No NLU training data needed.
2. MCP sub-agents — Config-only MCPOpenAgent definitions that call MCP tools and set slots. No custom Python classes or prompt templates.
3. Direct MCP tool calls — The appointment booking step calls book_appointment directly from a flow using call: + mcp_server: + mapping: syntax, bypassing the need for agent reasoning.
4. Slot collection — Flows collect user input (from_llm slots) before calling sub-agents. The MCP server handles type casting.
5. Response rephrasing — Rasa's built-in rephrase NLG rewrites templated responses via GPT-4o so they sound natural and contextual.

Testing

# Run all E2E tests
uv run rasa test e2e tests/

# Run a single test file
uv run rasa test e2e tests/test_out_of_scope.yml

Learn More

• Rasa Pro Documentation
• Building AI Assistants with CALM
• MCP Sub-Agents