Claude Code Subagents, Slash Commands, and MCP Servers: A European Dev Team Evaluation Guide

TL;DR: Why this matters for European dev teams approving Claude Code subagents, slash commands, hooks, settings, and MCP servers.

Claude Code's extension surface has matured faster than most engineering leads anticipated. The official product language is not "plugins" as a single generic category. The practical surfaces are subagents, slash commands, hooks, settings, memory files such as CLAUDE.md, and MCP servers that expose tools, resources, and prompts. For a 20-person software team in Warsaw or Lyon, choosing the wrong extension pattern can quietly erode developer confidence in AI tooling within weeks. Why this matters: technical leaders need an approval path before extensions start reading repositories, calling third-party APIs, or changing team workflows.

Source basis: Anthropic's subagents, slash commands, hooks, settings, and MCP documentation.

Subagents vs. Slash Commands vs. MCP Servers

Subagents are specialized AI assistants with their own context window, system prompt, and optional tool access. Anthropic documents project subagents under .claude/agents/ and user subagents under ~/.claude/agents/. Use them when a workflow needs a repeatable role such as code review, test repair, or release-note drafting.

Slash commands are reusable prompts. Project commands live in the repository and personal commands live in the user's Claude configuration. MCP servers can also expose prompts as slash commands with names like /mcp__github__list_prs.

Hooks run commands before or after Claude Code tool events. They are useful for logging, policy checks, and blocking unsafe tool calls, but they execute deterministic commands and Anthropic explicitly places responsibility for hook behavior on the user.

MCP servers are separate processes that expose tools, resources, and prompts over the Model Context Protocol. An MCP server might expose repository metadata, issue trackers, databases, or internal APIs. The protocol boundary matters for EU teams: data leaving the local process to a remote MCP server may constitute a transfer under GDPR Article 44 if that server is outside the EEA.

For a practical starting point on configuring Claude Code for your team's environment, see the Claude Code CLAUDE.md configuration guide for engineering teams.

Five Extension Patterns European Dev Teams Should Evaluate

1. Project subagents for repeatable review work. A project subagent can define when it should be used and which tools it may access. This is the right pattern for code review, test triage, documentation checks, and release-note drafting.

2. Project slash commands for common prompts. Use slash commands for low-risk prompts that developers run frequently, such as /review, /explain-change, or /write-tests. They are simpler than MCP because they do not add a separate process or external data flow.

3. Hooks for policy checks. Use PreToolUse hooks to ask, allow, or deny specific tool calls, and PostToolUse hooks for logging or validation. This is useful for teams that need a record of shell commands or a guardrail against sensitive file access.

4. MCP servers for external systems. Use MCP when Claude Code needs structured access to tools or data outside the current repository. Start with read-only scopes and explicit permission rules. Do not treat "MCP server" as a trust label. Treat it as an integration boundary.

5. Settings and memory for team defaults. Use CLAUDE.md for project context and .claude/settings.json for shared settings such as permissions, hooks, and environment variables. Keep .claude/settings.local.json for uncommitted personal settings only.

For a broader view of how autonomous agent workflows connect these extensions, see Claude Code agent mode and autonomous workflows.

Production-Ready vs. Experimental: How to Classify What You Adopt

The line between production-ready and experimental is not set by Anthropic. It is set by your team's operational tolerance. A useful three-question test:

Is this a prompt pattern, a role pattern, or an external integration? Use slash commands for prompt patterns, subagents for role patterns, and MCP servers only when a separate tool or data source is genuinely needed.

Does it write or only read? Any extension pattern that writes to git, file systems, or external APIs carries higher risk than read-only context. Treat write-capable subagents, hooks, and MCP tools as requiring explicit developer confirmation during the first 30 days of team adoption.

Is it version-controlled and permission-scoped? Project subagents, slash commands, and settings should be committed to the repository where they are intended to be shared. MCP servers should be version-pinned where possible, permission-scoped, and reviewed like other integration code.

For a structured framework on evaluating AI coding tools across these dimensions, see how technical leaders should choose an AI coding agent.

GDPR Considerations Specific to Claude Code Extensions

Claude Code's core model inference runs through Anthropic's API. EU teams using the commercial API are covered by Anthropic's standard DPA, which includes GDPR-compliant data processing terms. The extension layer introduces additional processors.

Three points to check before deploying any MCP server or externally networked extension to a team:

First, identify every outbound network call the integration makes. An MCP server that calls a US-hosted API is adding a data processor outside the EEA. If any code it processes contains personal data, even in comments or variable names referencing real users, you need a valid transfer mechanism: Standard Contractual Clauses, an adequacy decision, or confirmed EEA hosting.

Second, check whether the server logs requests server-side. Some community MCP servers log tool calls for debugging. Request logging of developer queries that contain business logic may constitute processing of confidential information.

Third, maintain an internal registry. Document each subagent, slash command, hook, and MCP server your team uses, who approved it, what data flows through it, and what the legal basis is for any cross-border transfer. This is the AI tooling equivalent of a processor and tool register.

The MCP marketplace guide covers the broader landscape of available MCP servers and how to evaluate them for enterprise use.

A Practical Adoption Sequence for Teams of 10-50 Engineers

Start with project slash commands and read-only subagents for the first four weeks. Commit the commands and subagents to the repository. Measure developer adoption by counting how often the commands or subagents appear in session logs. If fewer than 60% of engineers use them in week three, investigate friction before expanding.

Add read-only MCP servers in week five only when there is a specific integration need. No write capabilities, no broad scopes, and no external data residency questions beyond what you have already cleared.

Introduce write-capable or externally networked MCP servers only after the team has a tool registry and a lightweight DPA review process. Even a one-page checklist reduces the risk of a server silently routing production code to an unreviewed third-party service.

If your team is at the evaluation stage and not yet sure which extensions fit your architecture, an AI readiness assessment can identify the highest-value starting points before you commit to an integration path.

FAQ

Are Claude Code subagents the same as MCP servers?

No. Subagents are Markdown-defined assistants with their own prompt, context window, and optional tool access. MCP servers are separate processes that expose tools, resources, and prompts over the Model Context Protocol. Subagents are simpler to audit and version-control; MCP servers are more powerful but introduce additional process boundaries and potential data flows.

Can European teams use MCP servers without a GDPR-compliant DPA?

Only if no personal data or confidential code passes through the server or its downstream APIs. In practice, many coding sessions involve business logic that qualifies as confidential, if not personal data. The safe approach is to review each server's data processing terms before team-wide deployment and record the review in your tool register.

What is the simplest way to evaluate whether a subagent or command is ready for production?

Check three things: it is committed to the repository, it does not make write operations without explicit developer confirmation, and it has been tested by at least two engineers over five real sessions. If all three are true, it is operationally ready. Legal readiness is a separate check involving data flow review.

Further Reading

• Claude Code CLAUDE.md Configuration Guide: How to configure project-level instructions and defaults for your engineering team.
• Claude Code Agent Mode and Autonomous Workflows: How to connect subagents, commands, hooks, and MCP servers into multi-step workflows.
• How Technical Leaders Should Choose an AI Coding Agent: A structured evaluation framework for AI coding tools across capability, cost, and compliance dimensions.
• MCP Marketplace Guide 2026: A curated overview of the Model Context Protocol server ecosystem and how to assess servers for enterprise use.

If your team is deciding which Claude Code extension patterns are safe enough for shared use, the AI Readiness Assessment gives you a practical starting point for workflow, data, and governance review.