vibops-mcp

The MCP server for VibOps — The AI Infrastructure Engine. From code to GPU in one conversation.

The problem

Getting an AI app from code to production on GPUs requires stitching together 9+ tools — git, Docker, CI/CD, Helm, kubectl, GPU monitoring, cost management, compliance, alerting. Each with its own API, dashboard, and cost model. No single interface spans the full pipeline.

The solution

`vibops-mcp` connects your AI assistant to VibOps — the engine that clones, builds, deploys, scales, monitors, fixes, and bills your apps and agents on any GPU, any cluster, any cloud. One `pip install`, 94 tools, one conversation.

• Ship — clone repos, build containers, deploy models, run Helm/kubectl, trigger pipelines, submit Slurm jobs
• Operate — scale deployments, manage VMs (Proxmox/XO/vSphere), detect and remediate GPU anomalies
• Observe — GPU utilisation, workload breakdown, MTTR, cost estimates, live K8s deployments
• Govern — AI Act compliance, SOC 2/RGPD reports, immutable audit chain, policy management
• FinOps — per-agent cost tracking, budget enforcement, chargeback, spend trends, waste analysis

Every operation goes through your VibOps instance and is recorded in the immutable audit log.

Installation

pip install git+https://github.com/VibOpsai/vibops-mcp.git

Configuration

You need two environment variables:

Claude Desktop

Add to `~/.config/claude/claude_desktop_config.json` (macOS: `~/Library/Application Support/Claude/claude_desktop_config.json`):

{
  "mcpServers": {
    "vibops": {
      "command": "vibops-mcp",
      "env": {
        "VIBOPS_URL": "https://vibops.example.com",
        "VIBOPS_TOKEN": "your-token-here"
      }
    }
  }
}

Cursor

Add to `.cursor/mcp.json` in your project root, or to the global config:

{
  "mcpServers": {
    "vibops": {
      "command": "vibops-mcp",
      "env": {
        "VIBOPS_URL": "https://vibops.example.com",
        "VIBOPS_TOKEN": "your-token-here"
      }
    }
  }
}

Claude Code (CLI)

claude mcp add vibops vibops-mcp \
  -e VIBOPS_URL=https://vibops.example.com \
  -e VIBOPS_TOKEN=your-token-here

Available tools

Observation (16 tools — read-only)

Actions (18 tools — write)

Configuration (3 tools)

Agent Infrastructure Control Plane (12 tools)

The missing layer between your AI agents and your GPU fleet. Works with any framework (n8n, LangChain, CrewAI, Dify) — just point to the VibOps LLM Proxy.

Governance & Compliance (21 tools)

GPU FinOps (4 tools)

LLM Inference Proxy

VibOps includes a transparent OpenAI-compatible proxy (port 8004) that sits between your AI agents and LLM inference servers (vLLM, Ollama, TGI). Every inference request is logged with agent attribution for FinOps.

Your agents point to the proxy instead of the LLM directly:

# Before
OPENAI_BASE_URL=http://vllm:8000/v1

# After
OPENAI_BASE_URL=http://vibops-proxy:8004/v1

Add a `X-VibOps-Agent-Id` header to attribute costs per agent:

curl -X POST http://vibops-proxy:8004/v1/chat/completions \
  -H "X-VibOps-Agent-Id: pricing-agent-v2" \
  -H "X-VibOps-Team: supply-chain" \
  -d '{"model": "mistral:7b", "messages": [...]}'

The proxy captures: agent ID, team, model, tokens, latency, GPU cost — visible in the console FinOps dashboard and queryable via `get_agent_usage`.

Example prompts

"Clone my repo and deploy it on the GPU cluster."
"Deploy llama3:8b on vibops-dev with 2 replicas."
"Scale the inference deployment to 4 replicas on prod-cluster."
"What's our GPU utilisation trend over the last 7 days?"
"Show me the cost breakdown per cluster this week."
"Which clusters have open critical GPU alerts?"
"Are there any open GPU anomalies right now?"
"Scan my infrastructure and show discovered services."
"What's our AI Act compliance score and which controls are non-compliant?"
"Generate a SOC 2 report for Q1 2026."
"Verify the audit chain hasn't been tampered with."
"Which agent costs the most in GPU this month?"
"Show me the inference cost breakdown for the pricing agent."
"Which agents depend on the claude-opus-4-6 model?"
"Create a machine identity for the pricing-agent with a 1-year expiry."
"Show me the spend trend for the last 7 days and flag any waste."

Contributing

See CONTRIBUTING.md. All contributions require a DCO sign-off (`git commit -s`).

License

MIT — free to use, modify, and distribute. See LICENSE.

Built on FastMCP and VibOps — The AI Infrastructure Engine.