I built a self-hosted MCP control layer for enterprise systems — permissions, audit, human-in-the-loop

freen cikaliloar — Tue, 13 Oct 2026 12:49:21 +0000

Full disclosure: I'm the author.

What it is: a self-hosted MCP control layer. It exposes your enterprise systems as MCP tools and puts permissions, audit and approval gates on top of them.

Why I built it: every time we let an agent near a client's real systems, the same three questions came up — what can it see, how does it reach our systems, and who approved what it did. Nobody had a good answer that wasn't hand-rolled per project.

What's in it:

Per-dataset access control — configurable who-sees-what, sensitive fields auto-masked
Plugin engine — generates connectors from templates instead of bespoke code per system
Full audit trail — every call logged and traceable
Human-in-the-loop — critical actions are reviewed before they execute
Docker Compose, runs fully on your intranet — data never leaves your network

One honest note: this is a China-origin project, so the built-in connectors are China-native (Feishu/Lark, WeCom, DingTalk). That doesn't limit the engine — it generates the rest from templates. If you need one that isn't there, open an issue and I'll ship it.

It's not an agent. It doesn't think or plan. It's the layer that lets any agent in over MCP.

Repo: https://github.com/apex-freen/fde-mcp-blade
Live demo: https://fde.agent-plat.com

Happy to answer anything about the architecture — that's what I'm here for.

DEV Community: freen cikaliloar

I built a self-hosted MCP control layer for enterprise systems — permissions, audit, human-in-the-loop