Building agentic AI for SMB ops. Shipping production multi-agent systems on Claude Code + Codex. Operator at Grupo CPZ (Brazilian insurance brokerage holding). Available for advisory.
→ Twitter: @taricanof
I run a non-tech Brazilian SMB where production AI agents handle daily back-office: ETL across dozens of insurance carriers, financial pipelines, mailbox triage, operational reporting. The system runs on Claude Code (architecture, frontend), Codex (implementation ), and Hermes (productivity, research) — orchestrated by ~40 written ADRs.
In parallel, I ship and maintain open-source Model Context Protocol (MCP) servers as the infrastructure that makes the above work.
Production-grade MCP servers for operator stacks:
mcp-onedrive-sharepoint⭐6 — Microsoft Graph: OneDrive + SharePoint, 33 tools, device-code auth, alsoodsCLImcp-outlook⭐1 — Outlook / Exchange via MS Graph, 40 tools, 178 passing testsmcp-whatsapp⭐2 — WhatsApp via Baileys (QR pairing, rate limited, circuit breaker, pt-BR templates)mcp-server-trello— Trello, 23 tools, 93 unit testsmcp-gmail-calendar— Gmail + Calendar, 35 tools, OAuth multi-accountmcp-hub— MCP aggregation gateway behind one tool surface
- Read-after-write is default — agent self-report doesn't close a task
- Source-of-truth design — CRM wins even when empty so pending cases stay visible
- Fail-closed before fail-open — pipeline halts when upstream signal is unavailable
- OAuth-only auth — no paid API keys in production
- Governance by ADR — material decisions go to a written record, not chat history
Production agentic AI in non-tech regulated SMBs. The boring auditable closure of a workflow: pull data, validate, publish, log, fail in a known way, notify a human only when intervention is genuinely needed.
Available for advisory and fractional engagements. DMs open on Twitter.
Business systems for Grupo CPZ (Competenza, Interpar, SegDigital, Ferga) live in private repos. What's public here is the infrastructure tooling I'd want regardless of the operator job.