ark-ai-mcp

ark-ai-mcp

Claude Code plugin providing MCP wrapper skills via mcporter CLI. Each skill replaces dozens of MCP tool definitions with a single on-demand invocation, keeping your context window lean.

Skills

Skill	Description
/mcp-github	40 tools for PRs, issues, repos, code search, releases (complements gh CLI)
/mcp-linear	Linear issues, projects, teams, comments, documents, cycles, milestones, attachments
/mcp-context7	Library/framework documentation lookup via Context7
/mcp-deepwiki	Public GitHub repo documentation and architecture queries
/mcp-playwright	Headless browser automation — navigate, interact, screenshot, scrape

Prerequisites

1. mcporter installed globally:

   brew install steipete/tap/mcporter
   

2. ~/.mcporter/mcporter.json configured with MCP server definitions for linear, github, context7, deepwiki, playwright

3. Auth setup per server:

   • Linear: OAuth via mcporter auth linear
   • GitHub: GITHUB_TOKEN env var
   • Context7: No auth required
   • DeepWiki: No auth required
   • Playwright: No auth required (runs locally via npx)

Installation

# From GitHub
claude /plugin marketplace add HelloWorldSungin/ark-ai-mcp
claude /plugin install ark-ai-mcp@ark-ai-mcp

# From local submodule
claude /plugin marketplace add ./external/ark-ai-mcp
claude /plugin install ark-ai-mcp@ark-ai-mcp

Then enable in ~/.claude/settings.json:

{
  "enabledPlugins": {
    "ark-ai-mcp@ark-ai-mcp": true
  }
}

Usage

In any Claude Code session, invoke skills by name:

/mcp-linear          # Then ask about Linear issues, create/update issues
/mcp-github          # Then search code, get file contents
/mcp-context7        # Then look up library docs
/mcp-deepwiki        # Then query public repo documentation
/mcp-playwright      # Then navigate pages, fill forms, take screenshots

Each skill provides the mcporter call syntax and common examples. Use --output json for machine-readable results.

Adding a New MCP Skill

A template is available at skills/_template/SKILL.md.template — copy it as your starting point.

Step 1: Register the MCP server in mcporter

Add an entry to ~/.mcporter/mcporter.json under mcpServers:

{
  "mcpServers": {
    "my-server": {
      "description": "What this server does.",
      "baseUrl": "https://mcp.example.com/mcp"
    }
  }
}

Common auth patterns:

// No auth (Context7, DeepWiki)
{ "baseUrl": "https://mcp.example.com/mcp" }

// Bearer token from env var (GitHub)
{ "baseUrl": "https://api.example.com/mcp/", "bearerTokenEnv": "MY_TOKEN" }

// OAuth flow (Linear)
{ "baseUrl": "https://mcp.example.com/mcp", "auth": "oauth" }

For OAuth servers, run mcporter auth <server-name> after adding the entry.

Step 2: Discover available tools

mcporter list my-server --all-parameters

Save the output — you'll use it to build the tool catalog and understand parameters.

Step 3: Create the skill file

cp skills/_template/SKILL.md.template skills/mcp-myserver/SKILL.md

Edit the file following the structure below.

Step 4: Write the skill content

A good skill has four parts, in order of importance:

1. Objective with decision framework

Don't just say what the skill does — explain when to use it vs alternatives. Claude needs to decide between gh CLI, this skill, Claude-in-Chrome, etc. A comparison table or bullet list works well.

2. Tool catalog as compact tables

Organize tools into logical categories (3-8 categories). Each tool gets one row:

| Tool | Key Parameters | Description |
|------|---------------|-------------|
| `tool_name` | `param1`, `param2` | One-line description |

For tools with a method parameter (like GitHub's pull_request_read), list all method values in the description — this is the most common source of missed capabilities.

3. Multi-step workflows with bash examples

Include 2-4 workflows showing how to chain tools for real tasks — not just isolated calls. These are the most valuable part of the skill because they teach Claude what a complete task looks like.

Bad: mcporter call github.get_file_contents ... (isolated call, no context) Good: "Review a PR" workflow showing get overview → get diff → check status → submit review

4. Tips section

Focus on:

• Critical parameters that change behavior (method enums, ref formats, query syntax)
• Pagination patterns (page + perPage and their limits)
• Auth requirements and how to verify they're configured
• Performance gotchas (cold start, token cost, rate limits)

Step 5: Test the skill

# Verify mcporter can reach the server
mcporter call my-server.some_tool --output json