QuoxMCP · run & scale · MCP bridge
The AI infrastructure engine.
The Model Context Protocol bridge that gives Claude native access to your entire infrastructure. 168 tools, 8 resources, 7 prompts, a full audit trail, and a one-click install via .mcpb.
In plain words
What it is, where it lives, when to reach for it
- What is it
- A Model Context Protocol server that gives Claude direct access to your QuoxCORE instance's tools.
- Where do I use it
- Inside Claude Desktop or Claude Code, on the machine where you run Claude.
- When would I use it
- When you want to ask Claude to do things on your Quox platform for you.
- How do I use it
- Download
quoxmcp.mcpbfrom the GitHub releases page and double click it; Claude Desktop installs it.
QuoxCORE is the free, self-hosted platform underneath this. What is QuoxCORE
Below: live requests crossing the bridge, one real tool at a time, from Claude to governed execution and back.
QUOXMCP · protocol adapter
AEE envelope · VOLT run event · AOCL layer trace · org_id + user_id attributed
Live demonstration
AI that can actually operate infrastructure.
Most AI assistants can only describe what to do, then wait for you to paste the output back. With QuoxMCP, Claude calls the tools itself through the Model Context Protocol, no regex matching, no copy-paste workflow. Watch a real diagnosis: three tool calls, three audit records.
native tool callsiterative: investigate, diagnose, actfull AOCL trail on every actionper-agent RBAC
Architecture
Zero tool logic lives here.
QuoxMCP is pure protocol translation: a STDIO MCP server speaking JSON-RPC 2.0 on one side and the collector tool API on the other. It fetches tool definitions from the QuoxCORE collector at startup, registers them, and proxies every call back for execution. Routing through the SSH bastion, RBAC enforcement, plugin licence checks and the entire audit trail stay exactly where they belong: in QuoxCORE.
The handshake
Speaks MCP properly, from initialize onwards.
QuoxMCP implements the full Model Context Protocol session lifecycle. The client opens with initialize, QuoxMCP answers with its capabilities and server info, the client confirms with notifications/initialized, and only then do tool listing and tool calls begin.
Tool definitions are fetched from the collector at startup and registered per agent, so what Claude sees in tools/list is exactly what that agent is authorised to use, nothing more.
- STDIO transport, JSON-RPC 2.0 framing
- Tools, resources and prompts declared as capabilities
- Tool set resolved per agent ID at startup
→ initialize protocolVersion · capabilities ← result serverInfo: quoxmcp 1.2.0 · capabilities: tools, resources, prompts → notifications/initialized → tools/list ← result 168 tools for agent "quox" → tools/call ssh_exec { host, command } ← result exit 0 · audited
The tool surface
168 tools, fetched live at startup.
The catalog below is the orchestrator's tool set as served by the collector's tools/list, resolved per agent and expanded automatically as you license new plugin packs.
Monitoring and alerts
17prometheus_query · alerts_manage · uptime_sla_report · grafana_dashboards · metrics_queryWorkflows
15create_workflow · execute_workflow · validate_workflow · import_n8n_workflow · list_executionsVirtualisation
14proxmox_vms · proxmox_vm_action · proxmox_snapshots_list · proxmox_ha_status · proxmox_backup_statusPlatform and apps
14skills_list · quoxchat_search_knowledge · action_add · db_query · codebase_query · core_healthSecurity
13security_audit · security_failed_logins · wazuh_alerts · tls_cert_info · security_session_revokeNetwork
11network_check · network_mtr · network_dns · network_bandwidth · network_tracerouteFleet and remote execution
10ssh_exec · fleet_health_matrix · fleet_execute · quoxagent_status · remote_sshMedia and capture
10gemini_image_generate · gemini_image_describe · screencap_capture · screencap_compareSystem
9system_logs · system_services · system_cron · system_uptime · system_disk_usageContainers
8docker_status · docker_logs · docker_stats · docker_prune · docker_inspectGovernance and evidence
7approve_request · deny_request · list_evidence · get_envelope_chain · query_envelopesMemory and context
5memory_save · memory_search · memory_update · entity_note · session_searchOrchestration
3delegate_to_agent · orchestrate_objective · get_objective_resultsComms
3email_send · email_read · send_notificationCounts from a live tools/list against the collector. Licensed plugin packs, Proxmox, MikroTik, memory and others, surface their tools here automatically, no restart, and each pack ships specialist agents scoped to its domain.
The real console
Not a mock-up. The admin page itself.
The catalog above is generated from this exact screen: live sessions, a searchable tool list and a one-click connection test, all inside QuoxCORE.
- 1Read the catalog counts up frontThe header states tool, resource and prompt totals as plain numbers, not marketing copy.
- 2Pick a personalised or generic bundleDownload a bundle scoped to your org, or grab the generic v1.2.0 release instead.
- 3Watch every live sessionActive sessions list their agent, host, last seen time and recent tool calls.
- 4Search the tool catalog168 registered tools are searchable by name or description right from the dashboard.
- 5Delegate and orchestrate from the listdelegate_to_agent and orchestrate_objective sit in the same catalog as every other governed tool.
- 6Test the connection on demandOne button fires a live check against the session without leaving the page.
Captured from the real QuoxMCP admin page, not a mockup.
Distributed agents
Every host becomes an AI agent.
QuoxMCP deploys alongside QuoxAgent on every managed host. Each server gets Claude with native tool access, calling back to the central QuoxCORE collector for execution, RBAC and audit.
One install command provisions Node.js, downloads the QuoxMCP bundle from the collector, generates the MCP config with the service key, and secures the credentials. No manual configuration on remote hosts.
- Single install script: Node.js + QuoxMCP + config
- Service key distributed over SSH, never baked into scripts
- Each host identified by its agent ID for RBAC
- Tools execute via the collector, no direct infra access from hosts
$ quox agent deploy nw-web-01 ▸ provisioning node.js 20 done ▸ fetching quoxmcp bundle done ▸ writing mcp config + service key done ▸ securing credentials (0600) done nw-web-01 registered · agent_id: nw-web-01
Agent RBAC
Right tools, right agent.
Each agent's tool set is resolved by the collector at startup. WARDEN gets security tools, HYPERION gets Proxmox, QUOX gets everything. Pick an agent to see the request QuoxMCP makes and what comes back.
GET /api/v1/tools/list?agent_id=quox → 168 tools
Orchestrator · full tool surface
The tool set is resolved per agent ID at startup. What an agent cannot use, it never sees.
Context layer
8 resources, 7 prompts.
Beyond tools, Claude gets read-only context and pre-built operational workflows, served by the collector alongside the tool catalog.
MCP resources.
Read-only infrastructure context, available to Claude without a tool call. Fetched at startup and cached with a 30-second TTL for live data.
MCP prompts.
Pre-built operational workflows Claude can invoke with a single command. Each prompt includes structured arguments and templated messages with conditional sections.
Production ready
Auth and resilience.
Service key auth
Every collector request carries an X-Service-Key header. No anonymous access to infrastructure tools.
Exponential backoff
Failed collector calls retry with exponential delays up to 8 seconds, and a 30-second timeout.
Per-tool logging
Every tool call, resource read and prompt invocation is logged with timing, agent ID and correlation context.
Graceful shutdown
SIGTERM and SIGINT handlers ensure clean disconnection. No orphaned connections or lost audit entries.
New in v1.2.0
Every Claude tool call is now fully attributed.
Before v1.2.0, tool calls from Claude Desktop bypassed the QuoxCORE audit pipeline entirely. They executed, but nothing was traceable. That changes with v1.2.0: every tool call now produces a complete audit record attributed to the org and user who initiated it.
- AEE envelope: structured event with actor, action, and resource
- VOLT run:
tool.call.executedevent with timing and result - AOCL layer trace:
aocl.tool.executechained under a stable session correlation ID - org_id + user_id attribution: required fields in the install dialog
- Visible in the QuoxCORE dashboard audit viewer in real time
{
"aee": {
"actor": { "org": "acme", "user": "alice" },
"action": "tool.call.executed",
"resource": "ssh_exec"
},
"volt": {
"event": "tool.call.executed",
"tool": "ssh_exec",
"duration_ms": 312
},
"aocl": {
"layer": "aocl.tool.execute",
"session_id": "ses_01j...",
"correlation_id": "cor_9f..."
}
}Install
Install QuoxMCP.
One-click install for Claude Desktop. No terminal, no JSON editing.
# 1. Download the bundle quoxmcp.mcpb · double-click to open # 2. Fill 4 fields in Claude Desktop's install dialog Collector URL http://your-collector:9848 Service Key sk-your-service-key Org ID your-org-id User ID your-user-id # 3. Restart Claude Desktop · done
# Clone and install git clone https://github.com/quoxai/quoxmcp.git cd quoxmcp && npm install # Set env vars and start QUOX_COLLECTOR_URL=http://... \ QUOX_SERVICE_KEY=sk-... \ QUOX_ORG_ID=... QUOX_USER_ID=... \ npm start
Straight answers
What the audit trail actually proves today.
The bridge itself is real: the STDIO/JSON-RPC handshake, the per-agent tool set resolved at startup, and the service-key auth on every collector call are running code, not a mock-up.
The "fully attributed" audit trail is beta rather than proven end to end: the collector does emit an AEE, VOLT and AOCL record for every tool call, but that emission is unit-tested rather than confirmed against the live store, and the older regex-matched tool path (only reached when MCP itself is not configured) does not yet write the same record.
The tool, resource and prompt counts and the production-hardening claims above are QuoxMCP's own self-report: this repository does not yet carry an independent status review of its own.
Ready when you are
One click. 168 tools. Full audit trail.
Download the .mcpb bundle, fill 4 fields in Claude Desktop, restart. Your AI gets native infrastructure access with every call attributed and traceable.
Go deeper