An IT Superagent for Claude Code: Managing Employees' Agents

⚡ The gist in 5 seconds - The gist: a specialized IT superagent built on MCP answers employees' agents' questions about infrastructure instead of the administrator. - Where it's available: the prototype runs on an internal server behind nginx, accessible externally only via VPN. - Limitation: it's a working prototype in a pilot, not a finished solution; the author describes personal experience, not a universal recipe. ### 🔍 What was found The author is the sole administrator for roughly 60 employees. When several employees were given Claude Code with preconfigured instructions and read access to the infrastructure, the agents initially ran on a Windows terminal server. But two or three people with several sessions were enough to eat up all the terminal's RAM; no VMs were provisioned, since everything had already moved to the web, and the agents relocated to employees' home machines via VPN. After the move, two problems emerged. Centralized management was gone: the administrator no longer had access to the md files and Claude configs on home computers. And an "admin middleman" problem arose: an employee's agent would write a message, the author would forward it to his own Claude, which would prepare a reply for the agent, and the author would send it back. To eliminate this, he built an IT superagent — a prototype of ~4,000 lines of Python with FastAPI and MCP, SQLite in WAL mode. To the agents, the superagent looks like an MCP server (streamable HTTP), connected with a single command using a personal token. Inside are two RAG sources: an administrative server catalog (~570 files, ~940 fragments, CPU indexing takes ~9 minutes, then incremental updates every 10 minutes) and a corporate knowledge base on pgvector. Permissions are filtered down to the fragment level: an employee won't get a piece of a server description they aren't allowed to see. ### 💡 Why it matters The main benefit is that manual message relaying between AIs disappears. The superagent centralizes management: rules for all connected agents live in a single markdown file on the superagent's side and are passed on connection and in every response. Change the file, restart the service — behavior changes for everyone, wherever they work. And thanks to the RAG bases, the superagent answers with facts about the specific infrastructure rather than "theory from the internet" — solving the typical problem where an AI gives a textbook-correct result that doesn't fit a real company. ### 🧩 Context Before the superagent, the author tried two approaches. First, he built agents for specific tasks, but each time nuances known only to the specific person were missed, and refinement dragged on. Then he moved to the Paperclip platform in the spirit of a "company of agents": an analyst agent managed by a "CEO" agent was supposed to help employees create agents. It didn't take off: creating an agent proved labor-intensive, and the platform itself was too opaque — agents would loop and wander off to solve tasks unrelated to the original one. Tokens were burned, and often there was no result. That's when Claude Code with personal agents came along, which led to the "middleman" problem.

🔗 Read on habr.com