6.77 · Enhancements với MCP servers — Extending Claude Code

Mở đầu: Claude Code + MCP = Extendable

MCP servers extend Claude Code capabilities via chuẩn protocol.

Add MCP server

CLI command:

claude mcp add [server-name] [command-to-start-server]

Example:

claude mcp add sentry npx @sentry/mcp-server
claude mcp add docs uv run /path/to/docs_server.py
claude mcp add github docker run -i github-mcp-server

Claude Code on startup connect to registered servers. Tools available trong sessions.

Custom MCP server example

Build server expose "convert doc → markdown":

from mcp.server.fastmcp import FastMCP
from pydantic import Field

mcp = FastMCP("DocsMCP")


@mcp.tool(
    name="document_path_to_markdown",
    description="Convert PDF, Word, or other doc to markdown format."
)
def convert_document(
    path: str = Field(description="Absolute path to document file")
) -> str:
    # Use python libs to convert
    import pypandoc
    return pypandoc.convert_file(path, "md")


if __name__ == "__main__":
    mcp.run(transport="stdio")

Register:

claude mcp add docs uv run docs_mcp_server.py

Now in Claude Code:

Convert /path/to/report.docx to markdown

Claude uses your custom tool → returns markdown. Chain with Write to save.

Popular MCP integrations

Install từ respective repos. Each extends Claude Code.

Dev workflow example

Bug fix workflow

1. Claude Code in repo
2. MCP: sentry-mcp connected
3. MCP: jira-mcp connected

> Fix the top error in Sentry this week

Claude:
a. Uses sentry-mcp list_errors → top issue
b. Reads related code files (built-in)
c. Analyzes, proposes fix
d. Edits code (built-in)
e. Runs tests (built-in Bash)
f. Uses jira-mcp update_ticket → mark resolved
g. Git commit, open PR

Claude Code orchestrate across multiple systems.

Workflow: Team standup

claude mcp add github ...
claude mcp add slack ...
claude mcp add gcal ...

> Generate my standup for today

Claude:
a. github-mcp → my PRs since yesterday
b. gcal-mcp → today's meetings
c. slack-mcp → blockers mentioned in #me channel
d. Generate markdown summary
e. slack-mcp post_message → send to #standup

Automate mundane async work.

Custom team server

Your org có internal tools? Wrap in MCP:

@mcp.tool()
def lookup_customer(customer_id: str):
    """Query internal CRM."""
    return db.query(customer_id)


@mcp.tool()
def ship_to_staging(service_name: str):
    """Trigger internal deployment."""
    return ci.trigger_deploy(service_name, env="staging")


@mcp.tool()
def open_incident(severity: str, description: str):
    """Open PagerDuty incident."""
    return pagerduty.create(severity=severity, desc=description)

Entire team use Claude Code với company tools.

Popular sources

• github.com/modelcontextprotocol/servers — official curated
• mcp.pulse.com — community registry
• Anthropic docs — integration guides

Check regularly — ecosystem growing fast.

Security considerations

MCP servers = arbitrary code execution access.

Best practices

1. Only trusted serversvet official/OSS before install
2. Sandbox sensitivedon't give prod access
3. Review server codeespecially custom
4. Permission scopingread-only vs write
5. Audit logstrack what agent does

Production security: dedicated channel vs casual use.

When NOT to add MCP server

Overkill for

• 1 simple task (inline tool suffices)
• Prototype / 1-time use
• No team benefit

Worth building

• Team-wide tool
• Reusable across projects
• Expose to external (as product)

Claude Desktop MCP config

Thay vì CLI, Claude Desktop có config file:

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "docs": {
      "command": "uv",
      "args": ["run", "/path/to/docs_mcp.py"]
    },
    "sentry": {
      "command": "npx",
      "args": ["@sentry/mcp-server"]
    }
  }
}

Restart Claude Desktop → servers connected. Use qua chat.

Example: Custom analytics MCP

@mcp.tool()
def get_metric(metric_name: str, time_range: str = "7d"):
    """Query Prometheus/DataDog metric."""
    return metrics_client.query(metric_name, time_range)


@mcp.tool()
def list_alerts(severity: str = "critical"):
    """List active alerts."""
    return alerts_api.list(severity=severity)


@mcp.tool()
def silence_alert(alert_id: str, duration_hours: int):
    """Silence alert temporarily."""
    return alerts_api.silence(alert_id, duration_hours)

On-call engineer uses Claude Desktop + this MCP:

What alerts active? Silence the db pool one for 1 hour while I fix.

AI on-call buddy.

Team adoption pattern

Phase 1 (Week 1)

• 1 dev tries Claude Code + few MCP servers
• Learn workflows

Phase 2 (Week 3)

• Team tries selected MCP servers (GitHub, Slack)
• Build 1-2 custom team tools

Phase 3 (Month 2)

• Team-wide adoption
• Custom MCP servers for team-specific tasks
• Shared config file trong repo

Phase 4 (Month 6)

• Process automation: standups, retros, code reviews
• Org-level MCP for company systems

Gradual, measure ROI each phase.

Anti-patterns

Install every MCP server

20 servers → tool bloat, slow Claude.

Fix: Only install used servers.

Custom MCP without docs

Team uses but can't fix or extend.

Fix: README, schema docs, tests.

No version control for config

Dev's Claude Desktop config local → team inconsistent.

Fix: Commit to repo. Sync across team.

Forget security

MCP server có DB write access exposed.

Fix: Principle of least privilege. Read-only default.

Áp dụng ngay

Bài tập 1: Install 1 MCP server (30 phút)

Visit modelcontextprotocol.io. Install filesystem or github MCP. Use via Claude Code.

Bài tập 2: Design team MCP (20 phút)

Cho your team, identify 3 tools worth building as MCP server. Sketch tool schemas.

Tóm tắt bài học