toolhive MCP — Anwendungsfälle, Installieren & Live-Demo

Warum nutzen

Hauptfunktionen

Container-isolated MCP servers (Docker / Kubernetes)
Curated, verified MCP registry with preset configs
One-click install into Claude Desktop, Cursor, VS Code
SSO, encrypted secrets, audit logging, OpenTelemetry
Desktop app, web, CLI, and Kubernetes Operator

Live-Demo

In der Praxis

toolhive.replay ▶ bereit

0/0

Installieren

Wählen Sie Ihren Client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "toolhive": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/stacklok/toolhive"
      ],
      "_inferred": true
    }
  }
}

Öffne Claude Desktop → Settings → Developer → Edit Config. Nach dem Speichern neu starten.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "toolhive": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/stacklok/toolhive"
      ],
      "_inferred": true
    }
  }
}

Cursor nutzt das gleiche mcpServers-Schema wie Claude Desktop. Projektkonfiguration schlägt die globale.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "toolhive": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/stacklok/toolhive"
      ],
      "_inferred": true
    }
  }
}

Klicken Sie auf das MCP-Servers-Symbol in der Cline-Seitenleiste, dann "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "toolhive": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/stacklok/toolhive"
      ],
      "_inferred": true
    }
  }
}

Gleiche Struktur wie Claude Desktop. Windsurf neu starten zum Übernehmen.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "toolhive",
      "command": "TODO",
      "args": [
        "See README: https://github.com/stacklok/toolhive"
      ]
    }
  ]
}

Continue nutzt ein Array von Serverobjekten statt einer Map.

~/.config/zed/settings.json

{
  "context_servers": {
    "toolhive": {
      "command": {
        "path": "TODO",
        "args": [
          "See README: https://github.com/stacklok/toolhive"
        ]
      }
    }
  }
}

In context_servers hinzufügen. Zed lädt beim Speichern neu.

claude mcp add toolhive -- TODO 'See README: https://github.com/stacklok/toolhive'

Einzeiler. Prüfen mit claude mcp list. Entfernen mit claude mcp remove.

Anwendungsfälle

Praxisnahe Nutzung: toolhive

How to install a community MCP without trusting random `npx -y` scripts

👤 Security-conscious engineers ⏱ ~10 min beginner

Wann einsetzen: You want the GitHub MCP running locally but don't want to give it unrestricted filesystem access.

Voraussetzungen

Docker Desktop — docker.com
ToolHive desktop app — stacklok.com/download

Ablauf

Open ToolHive, search the registry

Install 'github' from the ToolHive registry.✓ Kopiert

→ Container pulled, config prompt for GITHUB_TOKEN
Auto-configure your client

Click 'Connect to Claude Desktop'.✓ Kopiert

→ Claude Desktop config updated; restart required
Verify isolation

Confirm the container only has the env vars you set and no host filesystem mounts.✓ Kopiert

→ Inspect container config in the ToolHive UI

Ergebnis: A sandboxed MCP with least-privilege access to your machine.

Fallstricke

Filesystem MCPs need a mount — default no-mount is too tight — Add a scoped mount (e.g. ~/Projects only) rather than full-home

How to deploy MCPs on Kubernetes for a platform team

👤 Platform engineers ⏱ ~60 min advanced

Wann einsetzen: You want org-wide standardized MCP availability.

Voraussetzungen

Kubernetes cluster + cluster-admin — Existing EKS/GKE/AKS or kind for dev
ToolHive Operator CRDs installed — kubectl apply -f the Stacklok-provided manifests

Ablauf

Declare MCPServer resources

kubectl apply -f mcp-github.yaml — declaratively register a GitHub MCP with scoped secrets.✓ Kopiert

→ Pod running; Portal lists it
Expose to users via Portal

Engineers log into the Portal with SSO and one-click install into their client.✓ Kopiert

→ Per-user configs auto-generated

Ergebnis: Central catalog + per-user install; auditable.

Fallstricke

Stdio MCPs don't fit network-exposed Kubernetes well — Prefer Streamable HTTP MCPs for k8s; use the gateway for stdio ones

How to produce an audit trail of all MCP tool calls

👤 Compliance, security teams ⏱ ~30 min intermediate

Wann einsetzen: Your org needs 'who called what tool when' for SOC2 or internal policy.

Ablauf

Enable OpenTelemetry export

In ToolHive settings, set the OTLP endpoint to your collector.✓ Kopiert

→ Traces appear in your backend
Tag by user

Enable per-user identity propagation via SSO.✓ Kopiert

→ Each trace has actor id

Ergebnis: Per-user, per-tool audit log.

Kombinationen

Mit anderen MCPs für 10-fache Wirkung

toolhive + mcphub

Use ToolHive for isolation, MCPHub for multi-server routing

Run each MCP in its own ToolHive container and expose them through MCPHub groups.✓ Kopiert

toolhive + unla

Combine ToolHive-run MCPs with Unla's REST-to-MCP conversion

Register our Unla-converted internal API in ToolHive so it runs containerized with SSO.✓ Kopiert

Werkzeuge

Was dieses MCP bereitstellt

Werkzeug	Eingaben	Wann aufrufen	Kosten
(platform) registry-install	mcp_name	Install a verified MCP	free
(platform) gateway-proxy	Transport + auth policy	Front your MCPs with policy	free
(platform) runtime-isolate	Container spec	Sandboxed execution	free + infra cost

Kosten & Limits

Was der Betrieb kostet

API-Kontingent: None at platform layer
Tokens pro Aufruf: Minimal overhead
Kosten in €: Desktop + CLI + open-source Operator are free. Enterprise plans for support/compliance.
Tipp: Start with the free tier; upgrade only if you need dedicated SLAs or advanced SSO/RBAC.

Sicherheit

Rechte, Secrets, Reichweite

Minimale Scopes: Per-MCP: least privilege env vars and mounts SSO for team deployments

Credential-Speicherung: Encrypted secrets manager built into ToolHive

Datenabfluss: Determined by downstream MCPs; ToolHive adds telemetry if configured

Niemals gewähren: Don't auto-grant filesystem mounts beyond what each MCP needs

Running too many MCPs in a single container defeats the isolation; one per container.

Fehlerbehebung

Häufige Fehler und Lösungen

Container starts then exits immediately

Check the MCP's required env vars are set (e.g., GITHUB_TOKEN). View logs in ToolHive UI.

Claude Desktop can't find the server after install

Restart Claude Desktop — config changes only load on launch.

K8s Operator CRD install fails

Ensure cluster-admin and supported k8s version (>= 1.27).

Prüfen: kubectl version

OTLP traces missing

Collector endpoint must accept OTLP/HTTP or OTLP/gRPC; confirm protocol match.

Alternativen

toolhive vs. andere

Alternative	Wann stattdessen	Kompromiss
MCPHub	You want multi-server aggregation more than container isolation	Not container-runtime focused
Docker MCP Toolkit	You prefer Docker's own integrated tooling	Less curated registry

Mehr

Ressourcen

📖 Offizielle README auf GitHub lesen

🐙 Offene Issues ansehen

🔍 Alle 400+ MCP-Server und Skills durchsuchen