02 / Agent Infrastructure
- 3 → 1
- CRMs behind one interface
- 1 week
- From start to shipped
- 0
- Permissions copied or rebuilt
- 50+
- Engineers trained on MCP
The problem
Three CRM products, Propforce, Jarvis and Carforce, each held their own data behind their own permission model. A question that spanned them meant three interfaces and three mental models, and any AI layer on top would have to answer the same question every unified system faces: whose rules decide who sees what.
The approach
- 01
A dedicated MCP server per product, each reading directly from that product's RDS, Elasticsearch and existing API layer. No copied data and no sync process to keep in step.
- 02
A custom MCP client with artifact support, multi-turn memory and cross-session context, so one natural-language interface sits in front of all three.
- 03
Shipped end to end in one week, with Claude Code and Cursor as the daily build tools.
- 04
Wrote the company's Agentic AI architecture guidelines afterwards, and ran internal sessions on MCP, multi-agent orchestration and LLM integration for 50+ engineers.
The decision
Chose
Per-product RBAC. Every request is validated by the source CRM's own auth model.
Rejected
A centralised permission layer in front of all three products.
Never copy permissions, route through them. Each product's auth model stays authoritative, three products share one client with zero auth rebuilt, and access stays correct as each product evolves on its own schedule. A central copy would have been one more thing to keep in sync, and the first thing to drift.
Stack