Unclassified

ChatGPT Connector Registration for IntegrateWise

Tier C — Historical. This document is a point-in-time snapshot preserved for reference. For current architecture, see CANONICAL_PLATFORM_ARCHITECTURE.md.

Overview

IntegrateWise MCP Server is registered as a ChatGPT Connector allowing enterprise ChatGPT users to access:

  • spine.query — unified operational data
  • spine.get — entity details
  • memory.search — organizational knowledge
  • memory.propose — capture insights
  • spine.connected_sources — connection status

ChatGPT can then use its own connectors (Salesforce, Gmail, etc.) to reach external tools, with IntegrateWise in the middle as the continuity brain.

Architecture

ChatGPT Enterprise
  ├─ ChatGPT Connectors (Salesforce, Gmail, etc.)
  └─ IntegrateWise MCP Server (our connector)
       ├─ spine.query → D1 vault → Salesforce/Gmail data
       ├─ memory.search → vector DB → org knowledge
       └─ memory.propose → proposals table → triage bot

Result: ChatGPT can read unified data AND capture insights

Registration Steps

1. Create Connector Metadata

Go to ChatGPT Enterprise Settings → Connectors → Create Connector

Fill in:

Name: IntegrateWise Spine
Description: Unified operational truth — query sales data, accounts, memories

Category: Data Integration
Authentication: OAuth 2.0

API Endpoint: https://mcp.integratewise.ai
Auth Endpoint: https://mcp.integratewise.ai/oauth/authorize
Token Endpoint: https://mcp.integratewise.ai/oauth/token

Schema Endpoint: https://mcp.integratewise.ai/schema.json

2. Configure OAuth (if using token-based auth)

Option A: API Key Auth (simpler)

Authentication Type: API Key
Header Name: Authorization
Header Format: Bearer {token}

Users will paste their MCP token in ConnectorSettings.

Option B: OAuth (more secure)

Create OAuth endpoints in gateway:

POST /oauth/authorize
  ↓
Generate authorization code
↓ (user grants permission in ChatGPT)
  ↓
POST /oauth/token (code) → MCP token

3. Register Tools

ChatGPT will query https://mcp.integratewise.ai/schema.json to discover tools.

Create endpoint:

// services/mcp-connector/src/handlers/schema.ts
export async function getSchema(actor: SpineMcpActor) {
  return {
    tools: [
      {
        name: "spine.query",
        description: "Query unified operational data",
        inputSchema: { /* from spine-mcp-server.ts */ }
      },
      {
        name: "spine.get",
        description: "Get entity details"
        inputSchema: { /* ... */ }
      },
      // ... rest of tools
    ]
  }
}

4. Deploy MCP Server

The MCP server is already built in services/mcp-connector/src/spine-mcp-server.ts.

Deploy to CF Workers:

cd services/mcp-connector
npm run deploy
# → https://mcp.integratewise.ai

Ensure bindings:

  • CACHE_DB → D1 database
  • SPINE_ENCRYPTION_KEY → secrets
  • VECTORIZE → vector index
  • CACHE → KV namespace
  • MCP_SERVICE_SECRET → shared secret with gateway

5. Test in ChatGPT

  1. Connect Connector:

    • ChatGPT Settings → Connectors → IntegrateWise Spine
    • Authorize (OAuth) or paste API key
    • Test connection

  2. Test Tools:

    • Ask ChatGPT: "What accounts are in the Spine?"
    • ChatGPT calls spine.query with entity_type="account"
    • Returns list of accounts with status, health, MRR

  3. Query Examples:

    • "Find stalled deals in Salesforce"
    • "What patterns have we seen in renewals?"
    • "Summarize recent activities for Acme Corp"

User Flow (ChatGPT Enterprise)

1. User opens ChatGPT Enterprise
2. Selects "IntegrateWise Spine" connector
3. AuthenticatesOAuth) → gets MCP token
4. Token stored in ChatGPT session
5. User asks question (e.g., "Show me high-risk accounts")
6. ChatGPT calls spine.query with MCP token
   ↓ (validated in mcp-connector/src/middleware/auth.ts)
7. Query executes against encrypted vault
8. Results returned to ChatGPT
9. ChatGPT generates natural language response

Example Conversation

User: "What are my top 3 expansion opportunities this quarter?"

ChatGPT (using IntegrateWise):

Calling: spine.query({
  entity_type: "account",
  status: "active",
  health_trend: "improving",
  arr_over: 100000,
  limit: 3
})

Result: [
  { id: "acc-123", name: "Acme Corp", arr: 250000, health_score: 0.85 },
  { id: "acc-456", name: "TechCorp", arr: 180000, health_score: 0.78 },
  { id: "acc-789", name: "InnovateCo", arr: 150000, health_score: 0.82 }
]

ChatGPT Response:

Your top 3 expansion opportunities this quarter are:

  1. Acme Corp ($250k ARR, 85% health)
    • Healthy trend, receptive to upsells
    • Recommendation: Schedule EBR next week
  2. InnovateCo ($150k ARR, 82% health)
    • Rising health score indicates engagement
    • Recommendation: Present new feature set
  3. TechCorp ($180k ARR, 78% health)
    • Steady health, growth potential
    • Recommendation: Evaluate expansion product fit

Tier-Based Feature Access

Free Tier: (no ChatGPT access)

Starter Tier:

  • ✅ spine.query
  • ✅ spine.get
  • ✅ spine.summary
  • ✅ spine.connected_sources
  • ✅ memory.search (semantic)
  • ✅ memory.propose (to governance)

Pro Tier:

  • All starter features +
  • ✅ Advanced memory queries
  • ✅ Custom entity types
  • ✅ Report generation

Enterprise Tier:

  • All features +
  • ✅ Full audit logs
  • ✅ Custom connectors
  • ✅ Dedicated MCP instance

Security

  1. Token Validation:

    • 15-minute expiry
    • HMAC-SHA256 signed
    • Validates on every tool call
    • See services/mcp-connector/src/middleware/auth.ts

  2. Tenant Isolation:

    • All queries scoped to tenant_id
    • Hard wall: tenant_id mismatch → 403 Forbidden
    • No cross-tenant data leakage

  3. Role-Based Access:

    • Actor permissions derived from role + tier
    • Fine-grained checks in middleware/actor.ts
    • Tool access gated by permissions

  4. Encrypted Vault:

    • All data at rest encrypted (AES-256)
    • Decryption key in CF Secrets (not queryable)
    • Vault reads always decrypt before returning

Monitoring

Monitor via CF Analytics:

GET /api/v1/mcp/metrics
→ tool calls per hour
→ error rates
→ latency p50/p95
→ tenant activity

Add to dashboard:

// services/gateway/src/metrics.ts
export async function getMcpMetrics(tenantId: string, hoursBack: number = 24) {
  // Query CF Analytics for MCP calls
  // Return: call_count, error_count, avg_latency, tools_used
}

Troubleshooting

"Connector not authorized"

  • User didn't authenticate OAuth
  • Token expired (> 15 min)
  • Solution: Re-authorize in ChatGPT settings

"Entity type not configured"

  • Tenant doesn't have entity type enabled
  • Solution: Admin enables in Settings → Spine Config

"Connector not connected"

  • Tenant doesn't have data source enabled (e.g., Salesforce)
  • Tool tried to query disconnected source
  • Solution: Connect source in /connectors

High latency

  • Vault queries slow (encrypted field searches)
  • Solution: Pre-populate vectorize for memory searches

Integration with Onboarding

When user completes onboarding:

  1. Role + connectors stored in D1
  2. tenant_spine_config populated
  3. User can now authenticate ChatGPT connector
  4. ChatGPT sees only authorized entity types + connectors

Next: Auto-Approve Workflow

When ChatGPT calls memory.propose() to capture insight:

memory.propose({
  content_type: "insight",
  title: "Renewal Pattern: EBR → 85% renewal rate",
  body: "Accounts with quarterly EBRs renew at 85% vs 60% baseline",
  confidence: 0.92
})
  ↓
Triage Bot scores proposal
  ↓
Score ≥ 0.85 → auto-approve to org memory
  ↓
Next user asks: "What renewal patterns have we learned?"
  ↓
ChatGPT calls memory.search() → finds this insight

This closes the loop: ChatGPT captures knowledge, promotes to org memory, answers future questions.

Files Summary

Created

  • services/mcp-connector/src/middleware/auth.ts (209 lines) - Token validation
  • services/mcp-connector/src/middleware/actor.ts (327 lines) - Actor resolution

Existing (Already Built)

  • services/mcp-connector/src/spine-mcp-server.ts - MCP tool definitions
  • apps/web/lib/auth/connector-approval.ts - OAuth flow

To Create

  • services/mcp-connector/src/handlers/schema.ts - ChatGPT schema endpoint
  • services/gateway/src/oauth.ts - OAuth authorize/token endpoints
  • services/mcp-connector/src/handlers/proposal.ts - Enhanced memory.propose

Next Steps

  1. ✅ Deploy MCP server to CF Workers
  2. ⏳ Create schema endpoint (handlers/schema.ts)
  3. ⏳ Register in ChatGPT Connector UI
  4. ⏳ Test OAuth flow
  5. ⏳ Verify tenant isolation + permission gating
  6. ⏳ Load test (1000+ concurrent ChatGPT connections)
  7. ⏳ Enable memory.propose governance loop

Success Metrics

  • ChatGPT connects successfully (< 1% auth failures)
  • Queries execute < 500ms p50, < 2s p95
  • Tenant isolation enforced (0 cross-tenant leaks)
  • Memory proposals flow through triage bot (> 80% auto-approved)
  • User engagement: avg 5+ queries/user/day