MCP Profiles: one Claude per mission

The problem: one generic Claude for everything

By default, Claude Code loads every MCP configured in ~/.claude.json. That works for daily development, but it causes three concrete issues:

• Noise: 15+ MCP servers loaded when most are irrelevant for the current task
• Startup time: each MCP spawns a process, even if you never use it during the session
• Context confusion: Claude may reach for a design tool during a security audit, or suggest a dev tool when you're analyzing SEO

The solution: create specialized profiles that load only the tools relevant to a given mission.

The mechanism: .mcp.json + CLAUDE.md per project

Claude Code has a two-level configuration system:

1. Global (~/.claude.json): loaded in every session
2. Project (.mcp.json at the project root): loaded on top of global when you launch Claude from that directory

Combining these two levels with a CLAUDE.md that sets the role and mission-specific instructions gives you a specialized Claude per task.

Concrete example: 4 profiles in production

Here is a real-world configuration with 4 distinct profiles.

Profile 1: Fullstack dev (global)

The base profile, always active. Configured in ~/.claude.json:

{
  "mcpServers": {
    "sequential-thinking": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"] },
    "context7": { "command": "npx", "args": ["-y", "@upstash/context7-mcp"] },
    "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] },
    "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem"] },
    "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] },
    "supabase": { "command": "npx", "args": ["-y", "@supabase/mcp-server-supabase"] },
    "shadcn": { "type": "url", "url": "https://www.shadcn.io/api/mcp" },
    "figma": { "type": "url", "url": "https://mcp.figma.com/mcp" }
  }
}

This profile covers reasoning, docs, code, design and data. It handles 80% of daily work.

Profile 2: SEO audit

Directory: ~/projects/seo-audit/

.mcp.json:

{
  "mcpServers": {
    "seo-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "seo-mcp"]
    },
    "pagespeed": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "pagespeed-insights-mcp"],
      "env": { "GOOGLE_API_KEY": "your-key" }
    },
    "gsc": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-server-gsc"],
      "env": { "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/credentials.json" }
    },
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    },
    "firecrawl": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-server-firecrawl"],
      "env": { "FIRECRAWL_API_KEY": "your-key" }
    }
  }
}

Claude gets access to Google Search Console, PageSpeed Insights, page crawling and on-page SEO analysis. Combined with global MCPs (GitHub for repos, filesystem for configs), it covers a full technical SEO audit.

Profile 3: UI/UX audit

Directory: ~/projects/ui-audit/

.mcp.json:

{
  "mcpServers": {
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    },
    "accesslint": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@accesslint/mcp"]
    },
    "pagespeed": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "pagespeed-insights-mcp"],
      "env": { "GOOGLE_API_KEY": "your-key" }
    },
    "ui-expert": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@reallygood83/ui-expert-mcp"]
    },
    "seo-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "seo-mcp"]
    }
  }
}

The focus is on accessibility (AccessLint), real browser navigation (Playwright) and UX expertise. The SEO MCP is there to check semantic structure, not for a full SEO audit.

Profile 4: Security audit

Directory: ~/projects/sec-audit/

.mcp.json:

{
  "mcpServers": {
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    },
    "semgrep": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@semgrep/mcp-server-semgrep"],
      "env": { "SEMGREP_APP_TOKEN": "your-token" }
    },
    "snyk": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "snyk-mcp-server"],
      "env": { "SNYK_TOKEN": "your-token" }
    }
  }
}

Three tools are enough here: Semgrep for static code analysis (SAST), Snyk for vulnerable dependencies (SCA), and Playwright for dynamic testing. A security audit does not need design or SEO MCPs.

The specialized CLAUDE.md: giving a role

The .mcp.json loads the tools. The CLAUDE.md provides context and behavior.

For the security profile for example, the CLAUDE.md defines:

• The role: senior cybersecurity expert, Red/Blue/Purple Team
• The frameworks: OWASP Top 10, MITRE ATT&CK, CIS Benchmarks, CVSS v3.1
• The workflow: reconnaissance, static analysis, dynamic testing, report
• The output format: findings with severity, CWE, reproducible PoC, remediation
• The absolute rules: no destructive exploits, no false positives, always a concrete remediation

Without this CLAUDE.md, Claude would use Semgrep and Snyk generically. With it, Claude follows a structured audit methodology and produces professional reports.

Shell aliases: switch in one command

Last piece of the puzzle: aliases in your .bashrc (or .zshrc) to launch the right profile instantly:

# Dev profile: global profile, works anywhere
alias claude-dev='claude'

# SEO profile: loads SEO audit MCPs
alias claude-seo='cd ~/projects/seo-audit && claude'

# UX profile: loads UI/accessibility audit MCPs
alias claude-ui='cd ~/projects/ui-audit && claude'

# Security profile: loads cyber audit MCPs
alias claude-sec='cd ~/projects/sec-audit && claude'

After a source ~/.bashrc, type claude-sec and you get a security-specialized Claude with Semgrep, Snyk, Playwright and a pentester CLAUDE.md.

Step by step setup

mkdir -p ~/projects/seo-audit
mkdir -p ~/projects/ui-audit
mkdir -p ~/projects/sec-audit

# Example for the SEO profile
cat > ~/projects/seo-audit/.mcp.json << 'EOF'
{
  "mcpServers": {
    "seo-mcp": { "type": "stdio", "command": "npx", "args": ["-y", "seo-mcp"] },
    "playwright": { "type": "stdio", "command": "npx", "args": ["-y", "@playwright/mcp@latest"] }
  }
}
EOF

cat > ~/projects/seo-audit/CLAUDE.md << 'EOF'
# Claude Code - SEO Expert

## Role
Full technical SEO audit. On-page analysis, performance,
indexation and structured data.

## Workflow
1. Crawl the site (Playwright + Firecrawl)
2. PageSpeed analysis on key pages
3. Google Search Console verification
4. Report with priorities and recommendations

## Output format
Structured Markdown report with scores, screenshots and actions.
EOF

echo "alias claude-seo='cd ~/projects/seo-audit && claude'" >> ~/.bashrc
echo "alias claude-ui='cd ~/projects/ui-audit && claude'" >> ~/.bashrc
echo "alias claude-sec='cd ~/projects/sec-audit && claude'" >> ~/.bashrc
source ~/.bashrc

claude-seo
# In Claude, type /mcp to see active servers

Why it's worth it

Going further

A few ideas to expand your profiles:

• Data profile: PostgreSQL MCP, Supabase and a CLAUDE.md as a data analyst
• Writing profile: Firecrawl MCP (research), Memory (style guide) and a CLAUDE.md as a writer
• DevOps profile: GitHub MCP, Sentry, a CLAUDE.md with your team's runbooks
• Monitoring profile: Chrome DevTools MCP, Playwright and a CLAUDE.md as an SRE watching performance

The .mcp.json + CLAUDE.md + shell alias combination turns Claude Code into a modular toolkit. Each profile is a ready-to-use expert.