omniskill

OmniSkill

Unified skill infrastructure for AI agents in Go.

Overview

OmniSkill provides a common interface for defining, registering, and invoking AI agent capabilities across multiple execution environments:

• skill/ - Core Skill and Tool interfaces
• registry/ - Skill registration and discovery
• mcp/server/ - MCP server runtime with tools, prompts, resources
• mcp/client/ - MCP client for connecting to remote servers
• mcp/oauth2/ - OAuth 2.1 Authorization Server for authenticated MCP

Skills can be invoked via:

• Library mode - Direct in-process calls without protocol overhead
• MCP Server - Expose via Model Context Protocol (stdio, HTTP, SSE)
• MCP Client - Consume remote MCP servers as local skills

Installation

go get github.com/plexusone/omniskill

Quick Start

Define a Skill

package main

import (
    "context"
    "github.com/plexusone/omniskill/skill"
)

func main() {
    // Create a tool
    addTool := skill.NewTool("add", "Add two numbers",
        map[string]skill.Parameter{
            "a": {Type: "number", Required: true},
            "b": {Type: "number", Required: true},
        },
        func(ctx context.Context, params map[string]any) (any, error) {
            a := params["a"].(float64)
            b := params["b"].(float64)
            return map[string]any{"sum": a + b}, nil
        },
    )

    // Create a skill
    mathSkill := &skill.BaseSkill{
        SkillName:        "math",
        SkillDescription: "Mathematical operations",
        SkillTools:       []skill.Tool{addTool},
    }
}

Library Mode

Call tools directly without MCP overhead:

import (
    "github.com/plexusone/omniskill/mcp/server"
    "github.com/modelcontextprotocol/go-sdk/mcp"
)

rt := server.New(&mcp.Implementation{
    Name:    "calculator",
    Version: "1.0.0",
}, nil)

rt.RegisterSkill(mathSkill)

// Direct invocation - no JSON-RPC, no transport
result, err := rt.CallTool(ctx, "add", map[string]any{"a": 1.0, "b": 2.0})

MCP Server Mode

Expose skills via MCP for Claude Desktop or other clients:

// stdio (for Claude Desktop)
rt.ServeStdio(ctx)

// HTTP with SSE
rt.ServeHTTP(ctx, &server.HTTPServerOptions{Addr: ":8080"})

// With OAuth2 authentication (for ChatGPT.com)
rt.ServeHTTP(ctx, &server.HTTPServerOptions{
    Addr: ":8080",
    OAuth2: &server.OAuth2Options{
        Users: map[string]string{"admin": "password"},
    },
})

MCP Client Mode

Connect to remote MCP servers and use them as skills:

import (
    "os/exec"
    "github.com/plexusone/omniskill/mcp/client"
)

c := client.New("my-app", "1.0.0", nil)

// Connect to MCP server
cmd := exec.Command("npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp")
session, err := c.ConnectCommand(ctx, cmd)
defer session.Close()

// Wrap as skill
fsSkill := session.AsSkill(client.WithSkillName("filesystem"))

// Use like any local skill
for _, tool := range fsSkill.Tools() {
    fmt.Println(tool.Name())
}

Registry

Central skill registration and discovery:

import "github.com/plexusone/omniskill/registry"

reg := registry.New()
reg.Register(mathSkill)
reg.Register(fsSkill)

// Discover all tools
for _, tool := range reg.ListTools() {
    fmt.Printf("%s: %s\n", tool.Name(), tool.Description())
}

// Initialize all skills
reg.Init(ctx)
defer reg.Close()

Auto-Registration

Skills registered with the runtime can auto-register with a registry:

reg := registry.New()
rt := server.New(impl, &server.Options{
    Registry: reg,  // Enable auto-registration
})

rt.RegisterSkill(mathSkill)  // Also registers with reg

Package Structure

github.com/plexusone/omniskill
├── skill/       # Core Skill and Tool interfaces
├── registry/    # Skill registration and discovery
├── mcp/
│   ├── server/  # MCP server runtime
│   ├── client/  # MCP client for remote servers
│   └── oauth2/  # OAuth 2.1 authorization server
└── doc.go

Documentation

• Getting Started
• Concepts
• MCP Server
• MCP Client
• API Reference

Feature Comparison

Design Philosophy

1. Define Once, Use Everywhere - Skills work in library mode, as MCP servers, or wrapping MCP clients
2. Protocol at the Edge - MCP is for external communication; internal calls bypass JSON-RPC
3. Type Safety - Generic handlers with automatic JSON schema inference
4. Composable - Skills can wrap other skills or remote MCP sessions

License

MIT License - see LICENSE file for details.