MCP und CLI integrieren

Bindet Agent-Fähigkeiten über typisierte Werkzeuge, Ressourcen und kontrollierte Unterprozesse an.

mcp-cli-integration

Teammitglieder mit diesem Skill

AgentFreya

Entdecke, wie der Skill hilft

Wann ist der Skill hilfreich?

Nutze diesen Skill für neue MCP-Server oder Kommandozeilenintegrationen.

Wobei hilft er dir?

Trennt Werkzeugaktionen, adressierbare Daten und wiederverwendbare Prompts.
Schützt Protokollausgaben und begrenzt Laufzeit, Datenmenge und Arbeitsbereich.
Regelt Serveridentität, lokale Erreichbarkeit und Prozessbeendigung.

Das könntest du fragen

Integriere einen lokalen MCP-Server mit sauberem JSON-RPC, begrenzten Antworten und zuverlässigem Abbruch hängender Aufrufe. Brich ein absichtlich hängendes Werkzeug ab und starte den Client neu. Es bleiben keine verwaisten Prozesse; Diagnosetext beschädigt den Protokollstrom nicht.

Ausführliche Dokumentation

Hier findest du die ursprüngliche Anleitung, Beispiele und Prüfschritte des Skills in ihrer Originalsprache.

Details…Weniger anzeigen

MCP & CLI Integration

Standardize external tool, resource, and subprocess integrations so the agent runtime interacts with third-party capabilities through clean lifecycle management, strict transport framing, and bounded security permissions.

Model Context Protocol (MCP) Architecture

MCP separates capabilities into three distinct primitives—use each for its intended purpose:

Primitive	Direction & Purpose	Design Rules
Tools (`tools/list`, `tools/call`)	Model-controlled actions with typed JSON Schema inputs (`inputSchema`) and optional `outputSchema`	Annotate with `readOnlyHint`, `destructiveHint`, `idempotentHint`, `openWorldHint`. Return operational failures inside the result with `isError: true` (and `structuredContent` when `outputSchema` is defined) rather than JSON-RPC protocol errors.
Resources (`resources/list`, `resources/read`, `resources/subscribe`)	Application/context-addressed read-only data identified by URI (`file://`, `db://`, `git://`)	Use resources for large files, schemas, or live logs so tools can return lightweight resource links (`resource_link`) instead of bloating tool outputs.
Prompts (`prompts/list`, `prompts/get`)	User- or workflow-invoked reusable templates with typed arguments	Keep prompt templates parameterized and versioned cleanly.

Transport & Protocol Discipline

stdio Transport (Local Subprocess Servers):
- Stdout purity: The MCP server must write only newline-delimited JSON-RPC 2.0 messages to stdout. Any console.log, library banner, or debug print on stdout corrupts the JSON-RPC stream immediately—route all diagnostic logging strictly to stderr.
- Subprocess lifecycle: Spawn with explicit working directory (cwd), minimal allowlisted environment variables (never dump the entire host process.env if it contains unrelated secrets), and clean shutdown (stdin.end() $\rightarrow$ SIGTERM $\rightarrow$ timeout $\rightarrow$ SIGKILL on the process group).
Streamable HTTP Transport (Remote / Shared Servers):
- Send JSON-RPC requests via POST to the MCP endpoint with MCP-Protocol-Version (and Mcp-Session-Id when stateful sessions are active).
- Handle both application/json single responses and text/event-stream (SSE) streaming responses, supporting resumption via Last-Event-ID when supported by the server.
- Validate Origin headers and bind local HTTP servers exclusively to 127.0.0.1 (never 0.0.0.0 without authentication) to prevent DNS-rebinding and cross-origin local execution attacks.

Defensive Client & Subprocess Guardrails

Treat third-party MCP servers as untrusted:
- Pin or verify tools/list schemas to prevent tool poisoning and rug-pull attacks (where a server silently mutates a tool's description mid-session to inject hidden instructions or exfiltrate files).
- Namespace tool names by server (<server>__<tool>) so one server cannot shadow built-in or high-trust tools.
Per-call timeouts and payload caps: Enforce a hard timeout (AbortSignal.timeout(ms)) and maximum byte limit on every tools/call and resources/read request so a hung MCP server or runaway CLI command cannot stall the entire agent worker.
Headless CLI Agent Integration: When invoking external agent CLIs (Claude Code, Codex, Gemini CLI), run non-interactively (stdin closed or structured protocol), stream structured JSONL events, capture exit codes and stderr ring buffers, and enforce workspace path sandboxing.

Enthaltene Dateien 1

SKILL.md