MCP Integration
Connect AI assistants to PodWarden using the Model Context Protocol
PodWarden includes a built-in MCP (Model Context Protocol) server that lets AI assistants manage your infrastructure through natural language.
What is MCP?
MCP is an open standard that lets AI applications (like Claude Desktop, Claude Code, and Cursor) call tools exposed by external services. PodWarden exposes 115 tools covering clusters, hosts, workloads, deployments, storage, ingress, secrets, and system configuration.
Two ways to connect
Direct connection
Your AI client connects directly to your PodWarden instance's /mcp endpoint using a pwm_ token. Best when your instance is reachable from wherever your AI client runs.
Hub proxy (recommended)
Your AI client connects to PodWarden Hub's MCP endpoint. Hub routes tool calls to your instance through a secure WebSocket tunnel. No ports to open, no VPN to set up.
Why use Hub proxy?
- No need to expose your PodWarden instance to the internet
- Works behind NAT, firewalls, and VPNs — the tunnel is outbound
- Manage multiple instances from a single MCP endpoint
- All tool calls are logged in PodWarden's audit trail
What you can do
With MCP connected, you can ask your AI assistant things like:
- "List all my clusters and their status"
- "Show me hosts with available GPUs"
- "Create a stack for llama.cpp with 2 GPUs"
- "Deploy the staging workload to the prod cluster"
- "What's wrong with the failing deployment?"
- "Show me container logs for the inference workload"
- "Find which cluster has the most free GPU capacity"
- "Rollback the last deployment"
Key features
| Feature | Description |
|---|---|
| Zero install | Built into the PodWarden container — no separate packages to install |
| 115 tools | Full infrastructure management: clusters, hosts, workloads, deployments, storage, secrets, config |
| Access levels | Choose read-only or full access — controls which tools are exposed |
| RBAC | Tokens have roles (viewer, operator, admin) — the AI can only do what the token allows |
| Audit logging | Every tool call through MCP is logged with token name, method, path, and status |
| Hub proxy | Connect through Hub — no ports to open, works behind NAT |
| Streamable HTTP | Uses the MCP Streamable HTTP transport — compatible with all modern MCP clients |
Getting started
Direct connection
- Set up a connection from your AI assistant
- Create a token with the appropriate role
- Start asking your AI assistant to manage your infrastructure
Hub proxy
- Enable the Hub tunnel in PodWarden Settings
- Create an API key in Hub Dashboard
- Connect your AI client to Hub's MCP endpoint
- Start managing your infrastructure