Add MCP server for AI agent integration

Objective

Add an MCP (Model Context Protocol) server that exposes OSAPI's system and network operations as MCP tools. This gives AI agents (Claude Code, Cursor, etc.) native tool access to OSAPI without HTTP boilerplate.

The MCP server will be a fourth server-type command alongside api server, agent, and nats server, reusing the same JobClient layer and Lifecycle pattern.

Design

• SDK: github.com/modelcontextprotocol/go-sdk (official Go MCP SDK)
• Transports: stdio (default, for local AI tools) + streamable HTTP (flag-selected, for remote agents)
• Tool scope: System + Network operations only (no job queue internals)
• CLI: osapi mcp server start [--transport http --port 3001]

Tools (5 total)

Each tool gets an optional target parameter (defaults to _any). Broadcast targets (_all, labels) return per-agent results.

Tool	JobClient Method	Parameters
system_hostname_get	QuerySystemHostname / Broadcast	target
system_status_get	QuerySystemStatus / Broadcast	target
network_dns_get	QueryNetworkDNS / Broadcast	target, interface
network_dns_update	ModifyNetworkDNS / Broadcast	target, interface, servers, search_domains
network_ping	QueryNetworkPing / Broadcast	target, address

Package structure

internal/mcp/
├── server.go      # MCP server implementing cmd.Lifecycle
├── tools.go       # Tool registration and handlers
└── types.go       # Input structs for tools

cmd/
├── mcp.go              # mcp command group
├── mcp_server.go       # mcp server subcommand group
└── mcp_server_start.go # mcp server start (NATS + JobClient + MCP wiring)

Config

Add to internal/config/types.go:

type MCP struct {
    Server MCPServer `mapstructure:"server,omitempty"`
}

type MCPServer struct {
    Port int `mapstructure:"port"`
}

Add to osapi.yaml:

mcp:
  server:
    port: 3001

Files to create/modify

File	Change
go.mod / go.sum	Add modelcontextprotocol/go-sdk
internal/config/types.go	Add MCP / MCPServer types to Config
osapi.yaml	Add mcp.server.port default
internal/mcp/server.go	New — MCP server with Lifecycle
internal/mcp/tools.go	New — tool registration + handlers
internal/mcp/types.go	New — input structs
cmd/mcp.go	New — mcp command group
cmd/mcp_server.go	New — mcp server subcommand
cmd/mcp_server_start.go	New — mcp server start command
internal/mcp/server_public_test.go	New — lifecycle tests
internal/mcp/tools_test.go	New — tool handler tests

Notes

• Tool handlers follow the same pattern as REST handlers in internal/api/system/ — check job.IsBroadcastTarget(), call the appropriate JobClient method, format results.
• stdio transport is the standard for local AI tool integration. The server runs as a subprocess spawned by the AI tool.
• Streamable HTTP is for remote deployments (replaces deprecated SSE).
• The NATS embedded server's Stop() doesn't need a context (internally bounded), so the adapter pattern from cmd/nats_server_start.go applies only to that server, not here.