Why MCPX

The Model Context Protocol gave AI systems a standard way to call tools — semantic code search, structured reasoning, databases, web browsing. Hundreds of MCP servers exist today.

But deploying MCP in production has three unsolved problems.

1. Context Cost

Loading MCP servers into an AI session is expensive. Each server dumps its full tool schema into the conversation context upfront. Five servers? That's 50,000-100,000 tokens consumed before the AI does anything.

Servers Loaded	Context Cost	Impact
3 servers	~30-60K tokens	Noticeable context reduction
5 servers	~50-100K tokens	Significant. Less room for code.
10 servers	~100K+ tokens	Context nearly full before work begins
Any count (mcpx)	0 tokens	Tools called on demand via Bash

The AI carries those definitions for the entire session — whether it uses them or not.

mcpx eliminates this. Instead of loading schemas, the AI calls mcpx <server> <tool> through Bash when it needs a tool. On-demand discovery, zero upfront cost.

2. No Security

Every MCP tool call is unrestricted. There's no authorization layer, no policy enforcement, no audit trail.

An AI agent connected to a Postgres MCP can DROP TABLE as easily as it can SELECT. A code search server can be pointed at files outside the project. A Slack MCP can post to any channel.

mcpx adds the missing security layer. Policy enforcement, security modes (read-only, editing), content inspection (SQL mutation blocking), argument validation (path traversal prevention), and a JSONL audit log that records every call.

3. Multi-Server Management

Each MCP server needs its own security profile — a database MCP needs different rules than a code search MCP. And when two developers work on different projects simultaneously, their daemons shouldn't interfere.

mcpx handles this automatically. Per-server security policies, scoped daemon isolation, and a single config file that manages everything.

The Insight

AI agents already know how to use terminals. They compose commands, pipe output, parse JSON. The terminal is the universal interface.

mcpx converts every MCP server into a CLI command — and adds the security controls that production teams need.

What This Unlocks

Context efficiency

Zero tokens upfront. The AI discovers tools with mcpx list and --help only when needed.

Security

Policy enforcement, audit logging, read-only modes. Teams can adopt MCP without giving AI agents unrestricted access.

Speed

Daemon mode keeps heavy servers warm. Sub-millisecond startup for the CLI itself. <5ms for tool calls to a warm daemon.

Composability

Every MCP tool becomes a UNIX command. Pipe between servers, redirect output, compose with jq, grep, xargs.

Scalability

Adding a new server doesn't increase context cost. 5 servers or 50 — the AI pays the same: zero tokens upfront.

What MCPX Is Not

• Not a replacement for MCP. mcpx wraps MCP servers — it doesn't replace the protocol.
• Not an AI client. It doesn't make decisions. It's infrastructure between the agent and the server.
• Not only for humans. The primary user is the AI agent. Humans benefit from the security and audit capabilities.

Why Go

• Single binary. No runtime dependencies. Ship one file.
• Sub-millisecond startup. The CLI must be faster than the MCP server.
• Cross-platform. macOS, Linux, Windows from one codebase.
• Concurrency. Daemon mode, socket handling, and transport management need goroutines.