20 Feb 2026
Why We Built mcplexer: Stopping Cross-Client Data Leaks in AI Tooling
MCP has no concept of scope. If you work across multiple clients, your AI agent can reach every connected tool server regardless of which project you are in. mcplexer fixes that.
The problem
AI coding agents talk to tool servers via the Model Context Protocol. MCP is useful. It is also completely unscoped.
If you connect a GitHub server, a ClickUp server, and a Slack server, every project on your machine can reach all three. Working on Client A's repo? Your agent can still read Client B's Slack channels, create tasks in Client C's workspace, and push to Client D's GitHub org. Nothing in the protocol prevents it.
For anyone working across multiple clients — consultancies, agencies, freelancers — this is a confidentiality problem. Not a hypothetical one. A single misrouted tool call or an auto-completed context pulling the wrong project's data is a breach.
We looked for existing solutions. There were none. So we built one.
What mcplexer does
mcplexer is a local MCP router that uses your current working directory to determine which tool servers your AI agent can reach. The directory you are in defines the security boundary.
Key design decisions:
- Deny-first evaluation — every tool request starts denied; policies must explicitly grant access
- Directory inheritance — policies cascade from parent to child, overridable at any level
- Human-in-the-loop — configurable approval prompts for sensitive tools, even if the policy allows them
- Encrypted credentials — API keys scoped to directory trees, encrypted at rest
- Full audit trail — every routing decision logged with directory, timestamp, and outcome
The AI client connects to mcplexer instead of individual servers. Routing is transparent. No config changes in your editor or agent.
The stack
Go. Single binary. SQLite for local state and audit logs. Everything runs on your machine — no cloud service, no data leaving your environment. MIT licensed.
Source: github.com/revitteth/mcplexer
How Revitt can help
We built mcplexer because we needed it for our own multi-client work. If your team uses AI coding agents across different projects and you are concerned about data isolation, we can help you set it up or adapt it to your workflow.
Read the full case study for the technical detail, or get in touch.