This is my actual production setup at Smartcat, not a theoretical one. Model Context Protocol (MCP) is Anthropic's standard for connecting AI to external tools. Each data source gets an MCP server, Claude reads from all of them, you schedule agents on cron, and they post results to Slack. No custom integrations. No middleware. Just configuration. By the end of the setup, Claude is connected to your real data and your first agent is scheduled to run tomorrow morning.
The config file is the foundation
Your MCP servers live in one JSON file, claude_desktop_config.json. Each server entry names the command, the path to the server, and the environment variables it needs. The one rule that matters: reference tokens with ${VAR_NAME} and set them in your shell, never hardcode them. Add them to your .bashrc or .zshrc so they persist.
The eight data sources
The setup wires up Jira (tickets, sprint status, blockers), Slack (discussions and escalation signals), Zendesk (support tickets and sentiment), Salesforce via Databricks (deals, pipeline, customer health), Gong via Weaviate (call transcripts, searched semantically), Google Calendar (meeting context), Notion (specs and process docs), GitHub (commits, PRs, release branches), and Google Drive (strategic planning artifacts). Gong and Zendesk go through Weaviate because transcripts and tickets are too unstructured for SQL, and a vector database finds "what did customers say about pricing" even when the word "pricing" never appears. After adding each server, test it with one specific query. Vague queries fail; "show me unresolved PROD bugs" beats "what's broken?"
Scheduling and delivery
Agents run on cron. A small run-agent.sh wrapper runs the agent, captures output, and posts it to a Slack incoming webhook so your team sees the report without you forwarding it. A cron line like 0 7 * * 1-5 fires an agent at 7 AM every weekday. A few things bite people: Salesforce tokens expire after 60 minutes so refresh them hourly, Slack revokes webhooks unused for 30 days, and large Weaviate queries need date filters to stay fast.
Start small
Do not turn on all 18 agents at once. Rotate API tokens every 90 days, set up weekly log reviews to audit what agents accessed, and add one agent per week so your teammates can absorb them. Pick one thing to try this week: install the MCP server for a single tool, get one test query returning live data, then schedule the Daily Focus Agent to run tomorrow at 7 AM.