🚀 Introducing MCPX: A Gateway for Governing AI Agent Tool Usage

As more teams start experimenting with AI agents using MCP, one challenge keeps coming up - there's no clear way to govern how agents access tools, or understand what's happening when those tools are called.

That’s why we built MCPX - an open-source gateway that helps you add visibility, guardrails, and permissioning around MCP usage. Whether you're testing locally or building toward more complex workflows, MCPX gives you control over how agents interact with your tool ecosystem.

Check it out:
👉 MCPX on GitHub

---

🧱 Why We Built MCPX

We’re seeing real traction in teams using MCP to let agents call tools like Slack, GitHub, Gmail, internal APIs, and more. But the operational gaps are clear:

• Agents can call tools they shouldn’t
• No way to group or gate sensitive actions
• No built-in audit or usage tracking
• No policies for managing overuse or privilege boundaries

This lack of governance is called out directly by OWASP in LLM07 - Excessive Agency, where AI agents can do more than they should, often unintentionally.

MCPX gives teams a simple gateway to safely connect agents to tools - with access controls and observability built in.

---

🔐 Access Controls Are Now Live

We recently shipped a major feature: Access Controls that let you group and define permissions over tools - across services.

You can now define tool groups like this:

toolGroups:
  - name: "writes"
    services:
      slack:
        - "post_message"
        - "post_reaction"
      gmail:
        - "send_email"
        - "send_attachment"
      github: "*" # allow all tools from GitHub

  - name: "reads"
    services:
      slack:
        - "read_messages"
        - "read_comments"
      gmail:
        - "read_email"
        - "read_attachment"

  - name: "admin"
    services:
      slack:
        - "create_channel"
        - "delete_channel"

From there, you can assign specific LLMs, agents, or users to one or more of these groups - allowing fine-grained, flexible control over what tools are allowed and when.

More on how it works:
👉 Defining Access Controls for LLM Agents

---

🧪 Start Local, Scale as You Grow

MCPX is lightweight and easy to run locally. But as your usage grows, you can deploy it in production, plug into observability tooling, and layer in more advanced policies.

We share that evolution here:
👉 From Local Experimentation to Production-Grade Infrastructure

---

🛠️ Try It Out or Contribute

• ⭐ GitHub: github.com/TheLunarCompany/lunar/tree/main/mcpx
• Docs: docs.lunar.dev
• Community: Join our Discord (linked in the repo)

We’re learning alongside the community and would love your feedback or ideas. If you’re using MCP today - or just exploring how to bring AI agents closer to your systems - MCPX is a layer worth trying.

Comments

[yonantan]

Guess what!
I've got a fever, and the only prescription is more MCPX!