horizon-mcp-demo-extended
Extends the horizon-mcp-demo with a second MCP server wrapping the Horizon Data Predictive Model, demonstrating the Claudeception pattern — a single Claude agent connecting to two governed data systems simultaneously.
Built as research and reference material for the Horizon Data Partners white paper: The Governed Data Layer: Why AI Agents Fail Without One, and How to Build It.
What this demonstrates
The core pattern: Two MCP servers, one agent, zero hardcoded connections between them.
| Server | What it exposes | Tools |
|---|---|---|
Insurance Data (from horizon-mcp-demo) |
P&C insurance semantic layer — loss ratio, claim frequency, earned premium by segment | list_models, get_model_details, get_metric_definitions, query_data |
| Predictive Model (this repo) | Professional services pipeline prediction model — win probability, fees, margin, milestone timing | describe_inputs, describe_outputs, validate_payload, score_opportunity |
The agent connects to both servers simultaneously, discovers what each one exposes, and reasons about how to bridge them — including explicitly surfacing the schema mismatch between the two systems and the governance risks of any mapping it proposes.
The schema mapping problem (Option C)
The two systems use completely different schemas:
Insurance data: product_type (auto/homeowners), state (TX/CA/FL/NY)
Predictive model: ServiceLine (Audit/Tax/Advisory), ClientType (Business/Individual), Industry (Healthcare/Financial Services/Technology/Manufacturing), NewVsExisting (Existing/New), LeadSource (Referral/Competitive)
These fields do not map cleanly to each other. The demo provides a pre-specified mapping for some questions, but Question 3 deliberately asks the agent to reason about the mapping problem before applying it — surfacing the governance risks of cross-system field mapping in a way that documents rather than hides the assumption.
This is the real-world use case: someone has a predictive model and a data source with different schemas, and needs an agent to connect them intelligently.
Project structure
horizon-mcp-demo-extended/
├── data/
│ └── seed_predictive_model.py # Generates predictive_model.duckdb from scratch
├── mcp_server/
│ └── predictive_model_mcp_server.py # Model-side MCP server (4 tools)
├── scripts/
│ └── run_two_server_agent.py # Claudeception demo — two servers, one agent
├── logs/ # Created at runtime — token logs per run
├── docs/ # Additional documentation
├── requirements.txt
├── .gitignore
└── README.md
Prerequisites
- Python 3.12
horizon-mcp-democloned at the same directory level as this repo- The two-server agent looks for the insurance MCP server at
../horizon-mcp-demo/mcp_server/horizon_mcp_server.py
- The two-server agent looks for the insurance MCP server at
horizon-mcp-demofully built (dbt seed && dbt runcompleted)
Setup (Windows)
1. Clone this repo
git clone https://github.com/christianashworth/horizon-mcp-demo-extended.git
cd horizon-mcp-demo-extended
2. Create and activate virtual environment
py -3.12 -m venv .venv
.venv\Scripts\activate
3. Install dependencies
python -m pip install --upgrade pip
pip install -r requirements.txt
4. Generate the predictive model database
python data/seed_predictive_model.py
This generates data/predictive_model.duckdb — a DuckDB replica of the SQL Server model's trained segment estimates. Takes about 30 seconds.
5. Run the two-server demo
$env:ANTHROPIC_API_KEY = "your-api-key-here"
python scripts/run_two_server_agent.py
Demo questions
| # | Question type | Servers used |
|---|---|---|
| 1 | Model understanding — what does the predictive model need? | Predictive Model only |
| 2 | Governed data query — insurance loss ratios by segment | Insurance Data only |
| 3 | Schema mapping surfaced — reason about mapping insurance fields to model inputs before scoring | Both |
| 4 | Cross-server analysis — score all homeowners segments and combine with loss ratio data | Both |
| 5 | Governance reflection — what decisions need to be made before using this mapping in production? | Both |
Predictive model input contract
| Field | Type | Allowed values | Notes |
|---|---|---|---|
ServiceLine |
string | Audit, Tax, Advisory | Required. Highest priority — dropped last. |
ClientType |
string | Business, Individual | Required |
Industry |
string | Healthcare, Financial Services, Technology, Manufacturing | Optional (nullable for Individuals) |
NewVsExisting |
string | Existing, New | Required |
LeadSource |
string | Referral, Competitive | Required. Lowest priority — dropped first. |
Predictive model outputs
| Output | Description |
|---|---|
WinPct |
Probability of winning (0.0–1.0) |
NetFees |
Estimated net fees in USD if won |
MarginPct |
Estimated margin percentage |
DaysSellToStart |
Estimated days from contract signing to work start |
DaysStartTo50Pct |
Estimated days from work start to 50% completion |
DaysStartTo100Pct |
Estimated days from work start to 100% completion |
Notes
- The predictive model database (
data/predictive_model.duckdb) is excluded from version control — generated locally by the seed script. - The model MCP server opens DuckDB in read-only mode. No write operations are possible.
- Token usage is logged per question per server in
logs/two_server_run_<timestamp>.json. - The Claudeception pattern works with any MCP-compatible agent, not just Claude — the two servers are independent and do not communicate with each other directly.
