Pragma
An operations layer for Claude.
- Status
- Working prototype
- Year
- 2026
- My role
- Product and engineering
- Stack
- Node.js, Model Context Protocol, Zero dependency CLI
The problem
Operational data sits across infrastructure, delivery, support and revenue tools. Even when every number is available, deciding what deserves attention stays slow and mostly instinctive.
How I approached it
Pull read only signals into a shared metric catalogue, score them against stage aware bands, and turn exceptions into actions ranked by impact, confidence and effort. The same state is exposed to Claude through MCP tools, so a decision can move between a conversation and an interface without losing context.
What I built
- Zero dependency Node.js CLI and MCP server
- Connector system for read only business signals
- Rules and scoring engine with explicit confidence
- Interactive local dashboard for health, goals, blind spots and next actions
Where it stands
Working prototype with strictly separated live and demo modes, so plausible sample data can never be mistaken for evidence.
What it taught me
Giving an AI assistant a disciplined data layer beats giving it more context. Ranking is only trusted when the reasoning behind the rank is legible.