Why MCPX

The Problem

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 every one comes with a hidden cost.

Loading MCP servers into an AI session is ruinously expensive.

Each server dumps its full tool schema into the conversation context. Five MCP servers? That's 50,000-100,000 tokens consumed before the AI does anything. The AI carries that weight for the entire session — whether it uses those tools or not.

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

Servers also initialize slowly. Some take seconds to spin up. Every new conversation pays that startup cost again.

The AI is drowning in tool definitions instead of solving your problem.

The Insight

What if the AI didn't need to load MCP servers at all?

AI agents already know how to use the terminal. They run grep. They run git. They run curl. They compose commands with pipes. They read stdout.

Give MCP servers the same interface and the problem disappears.

The Solution

mcpx converts MCP servers into CLI commands. The AI discovers and calls tools through Bash — the same way it runs any other command.

# No schema loaded. No initialization. Just a shell command.
mcpx serena search_symbol --name "UserAuth"

On-demand instead of upfront:

• Native MCP: 100K tokens loaded immediately, carried forever
• mcpx: 0 tokens upfront, small per-call cost only when used

The AI explores tools lazily. It calls mcpx list serena only when it needs Serena. It calls --help only when it needs a specific tool's flags. Context stays clean. The AI stays fast.

What This Unlocks

Context efficiency

The AI's context window is for reasoning, not for carrying tool schemas. mcpx moves tool definitions out of context and into the filesystem where they belong.

Speed

Daemon mode keeps heavy servers warm between calls. The AI doesn't wait for LSP initialization every time it needs code search.

Composability

AI agents are excellent at chaining shell commands. Two MCP servers that know nothing about each other can be piped together:

mcpx serena find_symbol --name "PaymentService" --json \
  | mcpx sequential-thinking think --problem - --total_thoughts 5

Scalability

Add 20 MCP servers to a project. The AI pays zero tokens for the ones it doesn't use. With native MCP, 20 servers would consume the entire context window.

What MCPX Is Not

Not a replacement for MCP. MCP is the protocol. mcpx is a better delivery mechanism.

Not an AI client. It speaks MCP protocol but never calls an LLM.

Not just for humans. It's designed for AI agents first. Humans benefit too, but the primary user is the AI in your terminal.

Why Go

Single binary. Zero runtime dependency. Sub-millisecond startup.

When an AI agent runs mcpx in a Bash call, startup cost matters. Go starts in under 5ms. The binary ships as a single file. No Python. No Node. No runtime tax.