# GxP MDM MCP Server - Cypher Tools for Claude / ChatGPT / Cursor

This MCP server exposes your **Computerized System Inventory as a Knowledge Graph** via Cypher queries. Any Claude/ChatGPT agent can now query GxP master data without hallucinating.

## Architecture

```
Claude / ChatGPT / Cursor
        |
        | MCP (stdio)
        v
  mcp_server.py (14 tools)
        |
        | Cypher queries
        v
  Neo4j (or MOCK mode: NetworkX + JSON) <- your MDM golden record
```

### Why Cypher?

- **Blast radius** is a graph traversal: `MATCH (start)-[:SENDS_VIA*1..3]->(downstream)` — impossible in SQL
- **Data lineage** for ALCOA+ investigations
- **Regulation ground truth** — agent can only cite clauses returned by Cypher, prevents hallucination

## 14 Tools Exposed

| Tool | Cypher | What it does |
|------|--------|--------------|
| `cypher_query` | Custom | Safe read-only Cypher for exploration |
| `list_all_systems` | `MATCH (s:ComputerizedSystem) RETURN` | Full inventory - auditors ask this first |
| `get_system_details` | Full subgraph | System + functions + e-records + supplier + interfaces |
| `get_blast_radius` | `MATCH (start)-[:SENDS_VIA*1..$depth]->(downstream)` | **Killer app**: What downstream GxP systems are impacted by change? |
| `get_upstream_lineage` | Reverse traversal | Where does data come from? |
| `get_applicable_regulations` | `(s)-[:HAS_FUNCTION]->(f)-[:REGULATED_BY]->(reg)` | Anti-hallucination: only these clauses can be cited |
| `get_system_interfaces` | `(s)-[flow:SENDS_VIA]->(target)` | API/file/manual interfaces with GxP flag |
| `find_validation_gaps` | `WHERE validation_status <> 'Validated'` | Non-validated GxP Direct systems |
| `find_periodic_review_overdue` | `WHERE next_review < date()` | Overdue periodic reviews |
| `find_high_risk_functions` | `WHERE is_gxp_critical AND risk=High` | E-sig, batch release, potency |
| `get_supplier_risk` | `(s)-[:SUPPLIED_BY]->(sup)` | Supplier audit status, SOC2 |
| `get_data_lineage` | `MATCH (s)-[:GENERATES]->(e:ElectronicRecord)` | Lineage for a record type |
| `get_gxp_inventory` | `WHERE gxp_impact='Direct'` | GxP Direct inventory |
| `assess_change_impact` | Composite | **Orchestrates 3 Cypher queries** + generates impact assessment per GAMP 5 / CSA |

## Quick Start (No Neo4j needed - Mock Mode)

```bash
cd gxp_mdm_mcp_server
pip install -r requirements.txt

# Mock mode: uses JSON + NetworkX, no Neo4j required
python scripts/test_tools.py

# Should show:
# - List 4 systems
# - Veeva QMS details with downstream SAP
# - Blast radius: Veeva -> SAP
# - REJECT for audit trail purge
# - Minor for version upgrade
```

Mock mode is perfect for POC and Claude Desktop testing.

## Prod Mode with Neo4j

```bash
# .env - set Neo4j creds
cp .env.example .env
# Edit .env with your Neo4j URI

# Start Neo4j
docker-compose up -d neo4j

# Load sample data + schema
python scripts/load_sample_data.py

# Test with Neo4j
python scripts/test_tools.py

# Start API harness (optional)
uvicorn src.api_server:app --reload --port 8000
# http://localhost:8000/cypher/list_all_systems
# http://localhost:8000/cypher/blast_radius?system_id=SYS-VEEVA-QMS-001
```

## Claude Desktop Config

1. Find your Claude config: `~/Library/Application Support/Claude/claude_desktop_config.json` (Mac) or `%APPDATA%/Claude/claude_desktop_config.json` (Win)

2. Add (use absolute path):

```json
{
  "mcpServers": {
    "gxp-mdm-cypher": {
      "command": "python",
      "args": ["/absolute/path/to/gxp_mdm_mcp_server/src/mcp_server.py"],
      "env": {
        "NEO4J_URI": "",
        "NEO4J_USERNAME": "neo4j",
        "NEO4J_PASSWORD": "password"
      }
    }
  }
}
```

For mock mode, leave `NEO4J_URI` empty. For Neo4j, set to `bolt://localhost:7687`.

3. Restart Claude Desktop. You should see 14 tools under 🔌.

4. Try prompts:

```
List all GxP Direct systems in my inventory
> calls list_all_systems(gxp_impact="Direct")

What happens if I change Veeva QMS? Show blast radius
> calls get_blast_radius(system_id="SYS-VEEVA-QMS-001")

Assess this change: Enable audit trail purge after 7 years for Veeva QMS
> calls assess_change_impact -> should REJECT per 21CFR11.10(e)

Assess Veeva upgrade from 24R1 to 24R2 with no e-sig change
> calls assess_change_impact -> should be Minor per CSA low risk
```

## Cursor Config

See `config/cursor_config.json.example` - add to `.cursor/mcp.json`

## ChatGPT (with MCP support)

If using ChatGPT custom GPT with MCP, use `config/chatgpt_mcp_config.json` as reference. ChatGPT will call tools via stdio.

## Cypher Queries - Ground Truth

All queries in `src/cypher_tools.py`. Key ones:

**Blast radius (the moat):**
```cypher
MATCH (start:ComputerizedSystem {system_id: $system_id})
MATCH path = (start)-[:SENDS_VIA*1..$depth]->(downstream:ComputerizedSystem)
WHERE downstream.gxp_impact IN ['Direct', 'GxP Relevant']
RETURN downstream.system_id, length(path) as distance
```

**Regulation anti-hallucination:**
```cypher
MATCH (s:ComputerizedSystem {system_id: $system_id})
OPTIONAL MATCH (s)-[:HAS_FUNCTION]->(f)-[:REGULATED_BY]->(reg)
RETURN collect(DISTINCT reg) as regulations
```

Agent must ONLY cite clause_ids returned here.

## From POC to Production

1. Replace `data/*.json` with real Veeva Vault API + ServiceNow CMDB + Okta
2. Add write tools (with approval workflow) for updating validation_status
3. Add vector search tool for regulation RAG (embed GAMP 5 2nd Ed)
4. Add periodic review agent that calls `find_periodic_review_overdue` on schedule

You now own the layer every CSV agent must query.

## Troubleshooting

- `No module named mcp`: `pip install mcp`
- Claude doesn't see tools: Check absolute path in config, restart Claude, check logs `~/Library/Logs/Claude/mcp*.log`
- Neo4j connection fails: Falls back to mock mode automatically - check `NEO4J_URI`

Good luck cornering the market.