MCP with Claude Code

Disponible en français

The Problem

With hundreds of MCP servers and thousands of tools:

• Context window fills with tool definitions you never use
• Token waste on intermediate results
• No trust verification for servers

Solution: Docker MCP Gateway

Docker’s dynamic MCP approach solves all three:

1. Trust - Verified servers in Docker catalog
2. Context - Only loads tools you actually need
3. Discovery - Agents find and configure tools autonomously

Setup

# Update Docker Desktop (enable MCP toolkit in beta features)
# Connect servers via Docker Desktop → MCP Catalog

Claude Code connects to Docker, Docker manages all MCP servers.

How It Works

Instead of loading all tools upfront, Docker provides:

• mcp_find - Search catalog by name/description
• mcp_add - Connect a server
• mcp_remove - Disconnect a server

The agent discovers and loads only what it needs per session.

Code Mode

Agents write JavaScript tools that chain MCP calls. Use this by default when an agent needs to orchestrate a large API surface such as Cloudflare, GitHub, Linear, Stripe, or internal admin APIs.

Without code mode:
- Many endpoint/tool definitions live in context
- Each call returns full results to context
- Context fills quickly and agent cost rises with API size

With code mode:
- Agent gets a small search/execute surface
- Agent writes custom tool using only needed API calls
- Results saved to volume, not context
- Only summaries/answers returned to model

This is especially important for Cloudflare automation. Do not expose every Cloudflare operation as independent always-loaded tools when the agent only needs to deploy one Worker or inspect one D1 database. Keep broad API catalogs behind search/execute, broker scripts, or Code Mode.

Example: Search GitHub → Save to Notion

1. Agent creates "github-to-notion" tool
2. Tool searches repos with multiple keywords
3. Results written to Notion database
4. Model only sees "29 repos saved"

Benefits:

• Sandboxed execution (secure)
• State persistence (volumes)
• Minimal context usage

Chrome DevTools MCP

Direct browser inspection from Claude Code.

Setup

// claude_desktop_config.json or .mcp.json
{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["@anthropic-ai/chrome-devtools-mcp@latest"]
    }
  }
}

Launch Chrome

# Mac
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222

# Windows
chrome.exe --remote-debugging-port=9222

# Linux
google-chrome --remote-debugging-port=9222

Capabilities

Tool	Use
chrome_navigate	Go to URL
chrome_screenshot	Capture page
chrome_click	Click element
chrome_type	Type text
chrome_evaluate	Run JS
chrome_logs	Get console logs
chrome_network	View requests

Claude can:

• Debug console errors
• Inspect network requests
• Check localStorage/cookies
• Test UI interactions
• Verify DOM state

Cloudflare MCP Servers

Cloudflare MCP access should follow the same boundary as deploy tokens: resource-scoped where possible, read-only unless mutation is explicitly needed, and routed through a broker or CI for production changes.

Available Servers

Server	Use
mcp-server-cloudflare	Manage Workers, KV, R2, D1
@cloudflare/tanstack-ai	TanStack AI adapters for Workers AI and AI Gateway

Setup

{
  "mcpServers": {
    "cloudflare": {
      "command": "npx",
      "args": ["@cloudflare/mcp-server-cloudflare"],
      "env": {
        "CLOUDFLARE_API_TOKEN": "your-token",
        "CLOUDFLARE_ACCOUNT_ID": "your-account-id"
      }
    }
  }
}

Capabilities

Workers:

• List/create/delete Workers
• Deploy code
• View logs

KV:

• Create namespaces
• Get/put/delete keys
• List keys

R2:

• Create buckets
• Upload/download objects
• List objects

D1:

• Create databases
• Run queries
• List tables

Recommended Setup

For Fenod Stack projects:

{
  "mcpServers": {
    "docker": {
      "command": "docker",
      "args": ["mcp", "gateway"]
    },
    "chrome-devtools": {
      "command": "npx",
      "args": ["@anthropic-ai/chrome-devtools-mcp@latest"]
    },
    "cloudflare": {
      "command": "npx",
      "args": ["@cloudflare/mcp-server-cloudflare"],
      "env": {
        "CLOUDFLARE_API_TOKEN": "${CLOUDFLARE_API_TOKEN}",
        "CLOUDFLARE_ACCOUNT_ID": "${CLOUDFLARE_ACCOUNT_ID}"
      }
    }
  }
}

Tips

• Use Docker gateway for dynamic tool loading (saves tokens)
• Prefer Code Mode/search/execute for large APIs instead of loading every endpoint as a tool
• Chrome MCP for frontend debugging
• Cloudflare MCP for infrastructure management, with resource-scoped tokens only
• Chain tools with code mode for complex workflows
• Results to files/databases, not context
• Add budget/rate limits before letting agents run autonomous Cloudflare or AI Gateway workflows