Context

Each department wanted an assistant that answers from its own procedures and policies, cites its source and does not make up numbers. It also needed to look up live data such as revenue, exchange rates, GA4, Search Console and mail, without letting the model write its own SQL.

There was no dedicated server: everything had to run locally on each employee's machine inside Claude Code.

What I did

I was the lead developer (72% of non-merge commits) and handled overall integration; a colleague owned the core technical design and main code review.

• Wrote 5 Python MCP servers exposing 50 tools: GA4 (14), Search Console (14), revenue and exchange rates (13), mail (6) and product lookup (3).
• Created 12 of the department skills, plus a router command that sends a request to the right department's agent.
• Built internal login and department-level permissions that run inside the Claude Code CLI, with admin-only folders.
• Encrypted stored mail passwords with Windows DPAPI and restricted file permissions; blocked passwords typed into chat from being logged.
• Packaged a setup for employees who do not use Git, with a locked settings script and onboarding guides.

How it works

An Obsidian vault is the single source of truth for company knowledge, organized by department. Each department has its own agent definition, skill and persona. A request flows like this:

• The employee logs in; their department decides which skills and folders they can reach.
• The router picks the department skill, which first asks clarifying questions before answering.
• Answers come from vault documents and must label their source.
• For live data, the skill calls a read-only MCP tool (only sending mail writes anything). GA4 and Search Console properties are pinned through environment variables, so the model cannot query other properties.

Results

• 5 MCP servers, 50 tools, all read-only except sending mail.
• 12 department skills authored by me.
• 72% of the repository's non-merge commits.

What I'd do next

I also built a self-improvement loop for skills (harvest conversations, mine fixes, gate on held-out scores, admin approves). It has 81 passing unit tests but has not been run on real data yet, so I do not count it as a result. Running it on real usage, and measuring answer quality per department before and after, would be the next step.