MCP Gateway Patterns

Heavy MCP servers consume context tokens. Three isolation strategies exist.

Critical Knowledge

MCP Registration Methods

Method	File Location	Correct?
`claude mcp add --scope user`	`~/.claude.json`	✅
`claude mcp add --scope project`	`.mcp.json`	✅
Manual `~/.claude/mcp_servers.json`	Legacy file	❌ NOT READ

Correct Registration:

bash

claude mcp add --transport stdio --scope user <name> -- <command> [args...]

Agent MCP Access Rules

`tools:` Setting	MCP Access	Example
Empty/Omitted	✅ All tools	`tools:`
Explicit list (no MCP)	❌ NO access	`tools: ["Read", "Write"]`
Explicit list (with MCP)	✅ Listed only	`tools: ["Read", "mcp__serena__*"]`

MCP Tool Naming Convention

Registration Type	Tool Name Pattern
Plugin MCP	`mcp__plugin_<server>_<server>__<tool>`
User MCP	`mcp__<server>__<tool>`

Strategy Selection (Empirically Verified)

📌 Default Recommendation: Daemon (SSE)

Daemon pattern provides the best balance: fast startup (1-2s) + zero token overhead + state sharing. Only use alternatives when specific constraints apply.

Criteria	Daemon (SSE) ⭐	Agent Gateway	Subprocess
Startup latency	1-2s	~1s	30-60s
Token overhead	Zero	~350/tool	Zero
State sharing	✅ Yes	✅ Yes	❌ No
Setup effort	Medium	Low	Low

Decision Guide

code

Start with Daemon (default)
    │
    ├── Can't run background process? → Agent Gateway
    │       (serverless, restricted env)
    │
    ├── Air-gapped / offline only? → Subprocess
    │       (no network, rare use)
    │
    └── Simple project, token OK? → Agent Gateway
            (quick setup preferred)

When to Use Alternatives

Constraint	Alternative
Cannot run background daemon	Agent Gateway
Serverless environment (Lambda, Cloud Functions)	Agent Gateway
Offline/air-gapped environment	Subprocess
MCP used < 2 times per session	Subprocess
Quick prototype, token budget OK	Agent Gateway

Token Savings by MCP (Measured)

MCP	Tools	Overhead	Subprocess Benefit
Serena	29	~10,150	✅ High
Playwright	25	~6,250	✅ High
Greptile	12	~3,600	⚠️ Medium
Context7	2	~400	❌ Not worth it

Rule: If tools × 350 > 5,000, consider subprocess isolation.

2-Layer Protocol

Layer 1 - Intent (common across MCPs):

Intent	Effect	Use
QUERY	READ_ONLY	Search, lookup
ANALYZE	READ_ONLY	Interpret, impact
MODIFY	MUTATING	Change files
EXECUTE	EXTERNAL_EXEC	External calls

Layer 2 - Action: MCP-specific tool names

Gateway Template

yaml

---
name: {mcp}-gateway
tools:  # Empty = all tools including MCP
model: sonnet
---

Responsibilities:

•Context activation
•Pre-validation for MODIFY
•Atomic JSON responses

Common Issues

"MCP tools not visible"

•Cause: Session not restarted after claude mcp add
•Fix: Restart Claude Code

"Tool not found in subagent"

•Cause: Agent has explicit tools: list without MCP
•Fix: Empty tools: or add MCP tools to list

"Wrong tool name format"

•Cause: Mixing plugin (mcp__plugin_x_x__) and user (mcp__x__) formats
•Fix: Check claude mcp list for actual server name prefix

Daemon Quick Start