The problem
Enterprise support teams spend the first part of every ticket rebuilding context: who the customer is, what was tried before, and what’s actually broken. The signals that a customer is unhappy are spread across a helpdesk, monitoring, call recordings and chat.
What we built
- An MCP server that gives AI agents safe, permissioned access to 90+ tools across helpdesk, monitoring, call-recording and data systems. Every agent uses it as the integration layer.
- A shared Claude skills marketplace with tiered review, CI and authoring standards, so a whole org can contribute agent skills without chaos.
- Autonomous agents on GitHub Actions + Claude Code: a briefing on every new ticket, SLA-breach and sentiment alerts, stalled-ticket detection, and weekly reporting.
- A customer-facing AI support agent that resolves most conversations on its own, plus a daily agent that turns its gaps into documentation drafts.
- AI ticket triage whose writes are confidence-gated and deterministic: the model judges, and boring code acts.
Principles
- Guardrails before autonomy.
- Humans in the loop where it matters.
- Measure the cost of every run.
- Ship small, and often.
Client work: described, not shown.