✓ 23 internal tools exposed through one MCP server
✓ Content published in ~4 hours instead of 2–3 days
✓ Manual ops down from ~15 hours a week to ~2
Before → after
| Metric | Before | After |
|---|---|---|
| Tools exposed to the agent | 0 | 23 |
| Content publish turnaround | 2–3 days | ~4 hours |
| Weekly manual ops hours | ~15h | ~2h |
A founder-led agency ran on talent but was drowning in ops. Every deliverable, a blog post, a client email, a campaign tweak, passed through too many tools and too many manual steps.
The challenge
- Tools (WordPress, HubSpot, Google Ads, Notion) didn’t talk to each other.
- The founder was the integration layer, copying data between systems.
- AI assistants could write copy but couldn’t publish, update, or report.
- No single interface to operate the stack end-to-end.
What we built
- Custom MCP server exposed 23 internal tools: create/update WordPress posts, query HubSpot contacts, read Google Ads performance, append to Notion.
- Auth & permissions scoped each tool so the agent could only act where approved.
- Publishing workflow lets the agent draft in their voice, then publish with one human approval.
- Reporting tools let the agent pull campaign numbers and summarize them in Slack.
- Onboarding docs so any team member’s AI assistant can connect to the same server.
The results
Content that took two to three days to publish now ships in about four hours. Manual ops dropped from ~15 hours a week to ~2, and the founder finally has one place where the whole stack is controllable, by humans and by the AI.
The takeaway
MCP is the missing layer between “AI that suggests” and “AI that does.” Expose your real systems as tools and the assistant becomes an operator.
Stack MCPWordPressHubSpotGoogle AdsClaude