Claude Code Mods vs Skills vs MCP Servers
The definitive architectural comparison: when to employ client-side pre-execution lifecycle interceptors (Mods), prompt-time context steering (Skills), JSON-RPC tool providers (MCP), or local process triggers (Settings Hooks).
Enforcement & Safety. Intercept, rewrite, or block shell/filesystem actions deterministically before local execution.
Behavioral Steering. Inject project guidelines, testing doctrine, and style constraints into reasoning context.
Data & External APIs. Provide structured tools to read Postgres, Jira, Slack, or GitHub without shell pipe hacks.
Environment Automation. Run raw bash scripts synchronously before Claude Code starts or after session exit.
The 4 Extension Dimensions
| Criterion | Claude Code Mods | Skills (CLAUDE.md) | MCP Servers | Settings Hooks |
|---|---|---|---|---|
| Execution Boundary & Scope | Client-side Runtime Interceptor Executes inside Claude Code CLI host before/after tool dispatch. | LLM Context Window Appended into system prompt at start of conversation turns. | External Daemon / Subprocess Independent OS process communicating via stdio or SSE RPC. | Host Shell Subprocess Raw bash scripts fired during startup, tool run, or teardown. |
| Interception & Cancellation | Deterministic (AST Blocking) Can hard-abort commands, rewrite parameters, or inject security guardrails (<18ms). | Non-deterministic (Soft) Model guidance only; the LLM may hallucinate or bypass instructions. | Tool-level Only Can reject incoming arguments, but cannot stop other CLI tools or bash execution. | Exit-code Gate Non-zero exit can cancel command, but lacks contextual AST tool inspection. |
| Latency Overhead | < 18ms AST Overhead In-process JavaScript/TypeScript event hooks with memoized parsing. | 0ms Runtime RPC (Adds Tokens) Zero process delay, but consumes context window quota on every prompt. | 5ms - 50ms JSON-RPC Process IPC serialization and standard I/O stream latency. | 20ms - 200ms Shell Fork OS process creation overhead per execution cycle. |
| Sandbox & Host Access Authority | Scoped Permissions via plugin.json Declared fs.read, fs.write, network egress, and env access parameters. | No Direct Host Access Prompt instructions only; model must use CLI tools to interact with system. | Server-configured Scope Defined by the specific MCP server binary implementation. | Full Host Terminal Authority Runs with the active terminal user shell privileges without restriction. |
| Packaging & Distribution | npm / /plugin install specifier Standardized package with plugin.json manifest and hooks schema. | Markdown file (CLAUDE.md) Git-tracked text files in project root or ~/.claude/skills. | JSON Config (claude_desktop_config) Declared command, args, and env parameters in JSON configuration. | Shell Script / settings.json Custom command string defined in settings.json. |
Detailed Architecture Breakdown
1. Claude Code Mods: Deterministic Client Interceptors
Mods run synchronously in the Node.js event loop of the Claude Code CLI. When Claude determines a tool execution plan (such as executing bash or writing a file), the mod intercepts the call payload before OS process fork. It can inspect the command AST, rewrite arguments, prompt the human developer for confirmation, or abort execution immediately.
2. Skills (CLAUDE.md): Prompt-time Reasoning Directives
Skills are plain markdown instructions loaded into the conversation context window. They guide Claude on coding conventions, test execution patterns, and architectural guidelines. However, skills are non-deterministic: an LLM may ignore or misinterpret prompt instructions during edge cases.
3. MCP Servers: Standardized Structured Tool Daemons
Model Context Protocol (MCP) servers run as external processes providing typed JSON-RPC tools and resources (e.g. database query tools, GitHub API clients, Figma connectors). They do not intercept local CLI actions, but rather expand the tool repertoire available to the model.
Ready to explore verified Claude Code mods?
Browse our directory of heuristic-scanned mods for security, context visualization, and workflow automation.