opsmate

One command to give Claude Code full understanding of your infrastructure

opsmate launches MCP servers that connect Claude Code directly to Kubernetes, Docker, Prometheus, and your infra files.
30 DevOps tools. Zero copy-paste. Single binary.

Why?

You're debugging a production incident at 2 AM. You switch between kubectl, Prometheus dashboards, Docker logs, and ChatGPT — copy-pasting context back and forth. The AI doesn't know your cluster state. You waste precious time explaining what you're looking at.

opsmate fixes this. One command — and Claude Code can see your pods, query your metrics, read your logs, and lint your configs. It acts on live infrastructure, not stale copy-paste.

$ opsmate
  opsmate — DevOps AI Assistant

  MCP Servers:
  ✔ kubernetes (context: production)
  ✔ docker (readonly)
  ✔ prometheus (http://prometheus:9090)
  ✔ file-analyzer

  Launching TUI...

Then just ask:

> Why is pod nginx-7b5f9 crashing?

[k8s_get_pods] [k8s_get_pod_logs] [k8s_describe] [prom_query]

Root cause: The container exceeds its 128Mi memory limit (~240Mi actual).

Recommended fix:
  kubectl set resources deployment/nginx --limits=memory=512Mi

Terminal UI

opsmate launches a native terminal interface (TUI) built with bubbletea. Claude Code runs under the hood — you get streaming responses, tool-use indicators, and multi-turn conversations right in your terminal.

• Dark minimal interface (Tokyo Night palette)
• Real-time streaming from Claude Code
• Tool usage indicators ([k8s_get_pods], [prom_query], etc.)
• Multi-turn conversations with session persistence
• > prompt — type and press Enter
• Esc to cancel, Ctrl+C to exit

Install

# Go install (recommended)
go install github.com/paffin/opsmate/cmd/opsmate@latest

# Or download binary
curl -sSL https://raw.githubusercontent.com/paffin/opsmate/main/scripts/install.sh | bash

# Or build from source
git clone https://github.com/paffin/opsmate && cd opsmate && go build -o opsmate ./cmd/opsmate

Prerequisites: Claude Code installed and authenticated.

30 Tools Across 4 Domains

_{* Destructive operations require confirmation and can be disabled with --readonly}

Real-World Scenarios

> My nginx deployment keeps restarting, help me figure out why

Claude: Let me investigate...
[k8s_get_pods]     → Found pod nginx-7b5f9 in CrashLoopBackOff
[k8s_get_pod_logs] → OOMKilled detected
[k8s_describe]     → limits.memory: 128Mi
[prom_query]       → container_memory_usage avg: 240Mi

Root Cause: OOMKilled — container uses ~240Mi but limit is 128Mi

Recommended actions:
1. Increase memory limit: kubectl set resources deploy/nginx --limits=memory=512Mi
2. Check for memory leaks in nginx config

Apply fix #1? [y/N]

> Analyze the Dockerfile in ./app/

Claude: Let me review it...
[file_lint ./app/Dockerfile]

Found 5 issues:
  [!] CRITICAL: Running as root (no USER instruction)
  [~] WARNING:  Using latest tag (FROM node:latest)
  [~] WARNING:  No HEALTHCHECK defined
  [i] INFO:     Consider multi-stage build to reduce image size
  [i] INFO:     .dockerignore not found

Want me to generate a fixed version?

> Show me what's happening in production

Claude: Collecting cluster status...
[k8s_get_nodes]  → 3 nodes, all Ready
[k8s_top nodes]  → CPU: 45%, Memory: 62%
[k8s_get_pods]   → 47 pods, 2 not Running
[prom_alerts]    → 1 firing: HighMemoryUsage on node-2

Cluster Overview:
  Nodes:  3/3 healthy (CPU: 45%, Mem: 62%)
  Pods:   45/47 running (2 pending in staging)
  Alerts: 1 firing — HighMemoryUsage on node-2 (87%)

node-2 memory is at 87%. Want me to investigate which pods
are consuming the most memory?

How It Works

                    ┌──────────────────────┐
                    │    opsmate CLI       │
                    │   (single Go binary) │
                    └──────────┬───────────┘
                               │
                    ┌──────────┴───────────┐
                    │  TUI (bubbletea)     │
                    │  streams from Claude  │
                    └──────────┬───────────┘
                               │
               ┌───────────────┼───────────────┐
               ▼               ▼               ▼
        ┌──────────┐    ┌──────────┐    ┌──────────┐
        │ K8s MCP  │    │Docker MCP│    │ Prom MCP │  ...
        │ (stdio)  │    │ (stdio)  │    │ (stdio)  │
        └────┬─────┘    └────┬─────┘    └────┬─────┘
             ▼               ▼               ▼
        K8s Cluster     Docker Host     Prometheus

1. Reads config from ~/.opsmate/config.yaml
2. Generates .mcp.json pointing to opsmate mcp <server> subcommands
3. Injects DevOps system prompt via CLAUDE.md
4. Launches a native TUI that streams from claude -p --output-format stream-json
5. Each MCP server communicates over stdio — no ports, no API keys
6. Session IDs enable multi-turn conversations via --resume
7. On exit — clean up .mcp.json and CLAUDE.md markers

Safety First

Configuration

servers:
  kubernetes:
    enabled: true
    kubeconfig: ~/.kube/config
    context: ""                    # empty = current context
    namespaces: []                 # empty = all
    readonly: false

  docker:
    enabled: true
    host: unix:///var/run/docker.sock
    readonly: true                 # safe default

  prometheus:
    enabled: false                 # enable when needed
    url: http://localhost:9090

  files:
    enabled: true
    scan_paths: ["."]
    rulesets: [dockerfile, kubernetes, compose, terraform]

safety:
  confirm_destructive: true
  max_log_lines: 1000
  redact_secrets: true

claude:
  model: claude-sonnet-4-20250514
  # custom_prompt: |
  #   Focus on security best practices

vs. Alternatives

Roadmap

• Kubernetes MCP server (11 tools)
• Docker MCP server (8 tools)
• Prometheus MCP server (7 tools)
• File analyzer with lint rules (4 tools)
• Native TUI with streaming responses (bubbletea)
• Terraform MCP server (plan, apply, state)
• Ansible MCP server (playbook, inventory)
• opsmate doctor — diagnose environment issues
• opsmate init — interactive setup wizard
• Plugin system for custom MCP servers
• Helm chart for in-cluster deployment
• Grafana MCP server
• CI/CD pipeline MCP (GitHub Actions, GitLab CI)

Contributing

We love contributions! Whether it's a new MCP server, a lint rule, or a bug fix.

git clone https://github.com/paffin/opsmate
cd opsmate
go build ./...
go test ./...

Each MCP server is self-contained in mcp/<name>/ with three files: server.go, tools.go, handlers.go. See CONTRIBUTING.md for detailed guidelines.

License

MIT

_{Built with mcp-go and Claude Code}

If opsmate helps you during an incident, consider giving it a star