# vibops-mcp

[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
[![Python 3.11+](https://img.shields.io/badge/python-3.11+-blue.svg)](https://www.python.org/downloads/)
[![MCP Tools](https://img.shields.io/badge/MCP_tools-93-58a6ff.svg)](https://github.com/VibOpsai/vibops-mcp)
[![Tests](https://img.shields.io/badge/tests-4%2C900_passing-brightgreen.svg)]()

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`, 117 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/HPE VME), 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

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

## Configuration

You need two environment variables:

| Variable | Description |
|----------|-------------|
| `VIBOPS_URL` | Base URL of your VibOps instance, e.g. `https://vibops.example.com` |
| `VIBOPS_TOKEN` | API token — create one in VibOps → Settings → API Tokens |

## Claude Desktop

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

```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:

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

## Claude Code (CLI)

```bash
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)

| Tool | Description |
|------|-------------|
| `list_clusters` | List clusters and GPU utilisation |
| `list_kubectl_contexts` | List available kubectl contexts |
| `get_cluster_deployments` | Live K8s deployment status for a cluster |
| `get_cluster_rate` | Get configured GPU cost rate for a cluster |
| `list_jobs` | List recent jobs with optional filters |
| `get_job` | Get job details and result |
| `get_job_metrics` | Job success rate, latency P50/P95/P99, error breakdown |
| `get_gpu_metrics` | Hourly GPU utilisation time-series |
| `get_workload_breakdown` | Job count by workload type |
| `get_mttr` | Mean Time To Resolve GPU alerts |
| `get_cost_estimate` | Estimated GPU spend |
| `list_gateways` | List registered gateways and status |
| `list_alerts` | List GPU alerts (open or resolved) |
| `list_secrets` | List secrets (names only, never values) |
| `list_providers` | List configured AI/GPU cloud providers |
| `list_pipelines` | List automation pipelines |

### Actions (18 tools — write)

| Tool | Description |
|------|-------------|
| `scale_deployment` | Scale a K8s deployment replica count |
| `deploy_model` | Deploy an AI model onto a GPU cluster |
| `helm_upgrade` | Run helm upgrade --install |
| `helm_uninstall` | Uninstall a Helm release |
| `run_kubectl` | Run an arbitrary kubectl command |
| `git_clone` | Clone a git repository |
| `create_secret` | Store an encrypted secret |
| `trigger_pipeline` | Manually trigger an automation pipeline |
| `slurm_get_cluster_info` | Get Slurm cluster info and partition details |
| `slurm_list_jobs` | List Slurm jobs with optional filters |
| `slurm_get_job_status` | Get status of a specific Slurm job |
| `slurm_get_job_output` | Retrieve stdout/stderr of a completed Slurm job |
| `slurm_submit_job` | Submit a new Slurm job |
| `slurm_cancel_job` | Cancel a running or pending Slurm job |
| `registry_list_repos` | List container registry repositories |
| `registry_list_tags` | List tags for a container image |
| `registry_check_image` | Check image details (size, layers, created date) |
| `registry_delete_tag` | Delete a stale image tag (requires confirmed=True) |

### HPE VME / Morpheus (13 tools)

| Tool | Description |
|------|-------------|
| `vme_list_instances` | List all VMs managed by HPE VME |
| `vme_get_instance` | Get detailed VM status and configuration |
| `vme_list_servers` | List physical hosts with CPU/memory |
| `vme_list_clouds` | List configured clouds/zones (KVM, VMware) |
| `vme_start_instance` | Start a stopped instance |
| `vme_stop_instance` | Stop a running instance |
| `vme_restart_instance` | Restart an instance |
| `vme_create_snapshot` | Create a snapshot (pre-migration) |
| `vme_list_snapshots` | List snapshots for an instance |
| `vme_convert_image` | Convert disk format (vmdk → qcow2) — VMware migration |
| `vme_list_virtual_images` | List available images and templates |
| `vme_detect_vm_waste` | Detect stopped/idle VMs with recommendations |
| `vme_get_activity` | Recent audit log from Morpheus |

### Configuration (3 tools)

| Tool | Description |
|------|-------------|
| `set_cluster_rate` | Set GPU cost rate for a cluster (admin only) |
| `register_gateway` | Register a new gateway (returns one-time token) |
| `delete_gateway` | Revoke a gateway |

### 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.

| Tool | Description |
|------|-------------|
| **FinOps per agent** | |
| `get_agent_usage` | GPU cost per agent — tokens, requests, cost, GPU-hours. *"Which agent costs the most?"* |
| `get_agent_usage_detail` | Drill-down on one agent — daily breakdown, model distribution, cost trend |
| `get_agent_budget` | Current budget + MTD spend for an agent |
| `set_agent_budget` | Set monthly spend limit — soft alert at 80%, hard block at 100% (HTTP 429) |
| **Model access control** | |
| `get_agent_model_rules` | List model access rules — which agent can use which LLM |
| `update_agent_model_rule` | Create a rule: glob patterns, deny-first. *"RH agents → Mistral only"* |
| **Identity lifecycle** | |
| `list_agent_identities` | List machine identities for agents |
| `create_agent_identity` | Create a new machine identity (key shown once) |
| `rotate_agent_identity` | Rotate the key for an existing identity |
| `revoke_agent_identity` | Revoke an identity immediately |
| **Dependency graph** | |
| `get_agent_dependency_graph` | Full org-wide graph: agent→model, agent→connector, agent→sub-agent |
| `get_agent_dependencies` | Dependencies for one agent — impact analysis before migration |

### Governance & Compliance (21 tools)

| Tool | Description |
|------|-------------|
| `list_anomalies` | List GPU anomalies with optional cluster/status filter |
| `get_open_anomalies` | Get all currently open anomalies |
| `resolve_anomaly` | Mark an anomaly as resolved |
| `list_compliance_controls` | List compliance controls (filter by framework) |
| `get_compliance_score` | Get the compliance score for a framework (or all) |
| `update_compliance_control` | Update status, notes, or evidence URL for a control |
| `list_compliance_reports` | List generated compliance reports |
| `generate_compliance_report` | Generate a SOC 2, RGPD, or HIPAA report asynchronously |
| `get_compliance_report` | Poll/retrieve a generated compliance report |
| `list_audit_logs` | Query the immutable audit log with filters |
| `verify_audit_chain` | Verify HMAC-SHA256 integrity of the full audit chain |
| `get_policy` | Get the current organisation policy |
| `update_policy` | Replace the organisation policy (immediate effect) |
| `list_eval_rubrics` | List LLM-as-judge evaluation rubrics |
| `evaluate_job` | Trigger LLM-as-judge evaluation for a job |
| `get_job_evaluations` | Retrieve evaluation results for a job |
| `get_ldap_config` | Get LDAP / Active Directory configuration |
| `update_ldap_config` | Configure or enable/disable LDAP integration |
| `get_siem_config` | Get SIEM push export configuration |
| `update_siem_config` | Set Splunk/Datadog SIEM destination |
| `push_to_siem` | Export audit events to configured SIEM |

### GPU FinOps (4 tools)

| Tool | Description |
|------|-------------|
| `get_budget` | Get current GPU budget and consumed spend |
| `get_chargeback` | Get chargeback breakdown by tenant for a given month |
| `get_spend_trend` | Get daily GPU spend trend (default: last 30 days) |
| `get_waste_analysis` | Identify idle GPU resources and cost optimisation opportunities |

## 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:

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

## License

MIT — free to use, modify, and distribute. See [LICENSE](LICENSE).

Built on [FastMCP](https://github.com/jlowin/fastmcp) and [VibOps](https://vibops.ai) — The AI Infrastructure Engine.