Extra Skills Tools

Self-hosted MCP server with 83+ tools for AI workflows — featuring 80%+ token reduction through 5 optimization layers.

GitHub stars License: MIT

---

Table of Contents

• Overview
• Quick Start
• Token Optimization (80%+ Savings)
• Claude Code Optimization
• All Tools (100+)
• Architecture
• Essential MCP Servers
• Docker Services
• Installation
• Configuration
• Troubleshooting

---

Overview

This MCP server provides 83+ tools for AI-powered development workflows, featuring:

• 🌐 Web Search & Scraping — SearXNG, Firecrawl, structured extraction
• 🐙 GitHub Integration — Repos, issues, commits, search
• 📺 YouTube — Transcripts, search, summarization
• 💻 Code Execution — Python sandbox, testing
• 📊 Data & Charts — Pandas, matplotlib visualization
• 🧠 Engineering Intelligence — Task classification, bug tracing, memory
• ⚡ Token Optimization — Built-in 80% token reduction

No API keys required for core functionality.

---

Quick Start

# 1. Clone repository
git clone https://github.com/simpletoolsindia/extra_skills_mcp_tools.git
cd extra_skills_mcp_tools

# 2. Run installation script
./install-claude-code.sh

# 3. Restart Claude Code
claude

# 4. Verify installation
docker compose -f docker-compose.local.yml ps

---

Token Optimization (80% Savings)

This server implements 5 layers of token optimization to maximize your context window:

Optimization Layers

Layer	Reduction	Description
Tool Trimming	80%	90 → 64 tools with concise descriptions
Web Content	80-97%	Clean markdown, strip nav/ads/scripts
Context Mode	98%	External SQLite storage for outputs
Lazy Loading	91%	Load schemas on-demand
Semantic Search	91%	Natural language tool discovery

Token Comparison

Metric	Before	After	Savings
Tool List	~13,500 tokens	~2,700 tokens	80%
Web Fetch	~8,000 tokens	~2,000 tokens	75%
Tool Output	~5,000 tokens	~50 tokens	98%
Full Workflow	~200,000 tokens	~30,000 tokens	85%

Optimization Tools

Tool	Usage	Tokens
quick_fetch	Ultra-fast title + summary	~25-300
fetch_web_content	Clean markdown with tracking	~2,000
fetch_structured	Article/product/table	~1,500
ctx_store_output	Store output externally	~50
tools_minimal	List without full schemas	~2,000
semantic_search	"search the web" → searxng_search	~500

Usage Examples

# Before: Fetch raw HTML (~8000 tokens)
fetch("https://example.com")  # Returns bloated HTML

# After: Optimized fetch (~300 tokens)
quick_fetch(url="https://example.com", max_tokens=1500)
# Returns: {title: "Example", summary: "...", tokens: 300}

# Store large outputs externally (98% reduction)
ctx_store_output(
    tool_name="github_repo",
    arguments={"owner": "anthropics"},
    output={"repo": "claude-code", "stars": 15000}
)
# Returns: {"ref": "@ctx:default:abc123", "size_bytes": 54}
# Instead of storing 500+ tokens, just store the reference

---

Claude Code Optimization

Model Selection Strategy

Use the right model for the right task:

Model	Best For	Cost	When to Use
Sonnet	Most coding	$3/1M tokens	Default choice, ~60% cheaper
Haiku	Code review, docs	$0.25/1M tokens	Routine tasks, fixes
Opus	Complex refactoring	$15/1M tokens	Architecture, deep debugging

Recommended Settings

Add to ~/.zshrc or ~/.bashrc:

# Model Settings
claude config set --claude-code-subagent-model sonnet
export HAIKU_MODEL=haiku

# Thinking Token Limit (~70% savings)
# Default: 32,000 tokens
# Recommended: 10,000 tokens
export MAX_THINKING_TOKENS=10000

# Compaction Settings (better performance)
# Default: 95% context before compact
# Recommended: 50% for more working room
export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50

# MCP Server Limit
# Keep under 10 MCP servers, 80 total tools
# More = reduced effective context window

Quick Commands

Command	Purpose
/cost	Monitor token usage and costs
/clear	Free context reset between tasks
/compact	Manual compaction at breakpoints
/context	Check current context usage

MCP Server Best Practices

Warning: Each MCP server adds tool definitions to context. Excessive servers reduce effective context from 200k to ~70k.

Recommended: Keep under 10 MCP servers, 80 total tools.

10 Strategies to Reduce MCP Token Bloat

1. Design tools with intent — Single purpose, clear inputs/outputs
2. Cache aggressively — Identical queries hit cache
3. Minimize server usage at runtime — Enable only when needed
4. Group tools by domain — Logical grouping reduces confusion
5. Deploy subagents — Route routine tasks to Haiku
6. Just-in-time context loading — Load schemas on-demand
7. Externalize computational results — Store large outputs
8. Apply advanced data filtering — Filter at extraction time
9. Externalize cross-cutting concerns — Centralize auth/errors
10. Keep tools lean — Runtime handles concerns centrally

Cost Comparison

Scenario	Before	After	Savings
1 hour coding	$2.50	$0.75	70%
1 day research	$8.00	$2.40	70%
1 week project	$35.00	$10.50	70%

---

All Tools (100+)

🌐 Web Search & Scraping (10)

Tool	Description
searxng_search	Web search via SearXNG (Pi5)
search_images	Image search
search_news	News search
searxng_health	Check SearXNG status
fetch_web_content	Clean markdown extraction
fetch_structured	Article/product/table extraction
quick_fetch	Ultra-fast title + summary
scrape_dynamic	JavaScript-heavy pages (Playwright)
firecrawl_scrape	Advanced scraping
webclaw_extract_article	Article extraction

🐙 GitHub (6)

Tool	Description
github_repo	Repository information
github_readme	README content
github_issues	List issues
github_commits	List commits
github_search_repos	Search repositories
github_file_content	Get file content

📺 YouTube (6)

Tool	Description
youtube_transcript	Get transcript
youtube_transcript_timed	Timestamped transcript
youtube_search	Search videos
youtube_video_info	Video metadata
youtube_batch_transcribe	Batch transcription
youtube_summarize	Summarize transcript

📰 Hacker News (6)

Tool	Description
hackernews_top	Top stories
hackernews_new	Newest stories
hackernews_best	Best stories
hackernews_ask	Ask HN
hackernews_show	Show HN
hackernews_get_comments	Get comments

💻 Code Execution (4)

Tool	Description
run_code	Sandboxed execution (Python/JS/Bash)
run_python_snippet	Python with common imports
test_code_snippet	Test code output

📊 Data & Charts (11)

Tool	Description
pandas_create	Create DataFrame
pandas_filter	Filter data
pandas_aggregate	Aggregate/group data
pandas_correlation	Compute correlation
pandas_outliers	Detect outliers
plot_line	Line chart
plot_bar	Bar chart
plot_pie	Pie chart
plot_scatter	Scatter plot
plot_histogram	Histogram
generate_chart_spec	Ant Design spec

🧠 Engineering Intelligence (17)

Tool	Description
engi_task_classify	Classify task type
engi_repo_scope_find	Find relevant files
engi_flow_summarize	Get execution flow
engi_bug_trace	Pinpoint bug causes
engi_implementation_plan	Generate implementation plan
engi_poc_plan	Scaffold POC
engi_impact_analyze	Estimate blast radius
engi_test_select	Select minimum tests
engi_doc_context_build	Build documentation
engi_doc_update_plan	Plan docs updates
engi_memory_checkpoint	Save task state
engi_memory_restore	Restore checkpoint
thinking_session_create	Create thinking session
thinking_step	Add reasoning step
thinking_summary	Get summary
analyze_problem	Structured analysis

⚡ Optimization Tools (14)

Tool	Description
get_token_stats	Token optimization stats
quick_fetch	Minimal token fetch
fetch_web_content	Optimized extraction
fetch_structured	Structured extraction
fetch_with_selectors	CSS selector extraction
ctx_store_output	Store output externally
ctx_get_output	Retrieve stored output
ctx_search	Search stored outputs
ctx_session_overview	Session summary
ctx_stats	Context mode stats
tools_minimal	Lazy tool list
tools_describe	Load schemas on-demand
tools_search	Search tools
semantic_search	Natural language search

💾 Files & HuggingFace (10)

Tool	Description
file_read	Read file
file_write	Write file
file_list	List directory
file_search	Search files
huggingface_search_models	Search models
huggingface_search_datasets	Search datasets
huggingface_model_info	Model info
huggingface_trending	Trending models
markitdown_html_to_md	HTML → Markdown
markitdown_url_to_md	URL → Markdown

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      Claude Code                                  │
│  ┌───────────────────────────────────────────────────────────┐  │
│  │  MCP Servers (< 10 recommended)                            │  │
│  │                                                            │  │
│  │  • mcp-server (83 tools) → Port 7710                    │  │
│  │  • github (10 tools) → NPX                               │  │
│  │  • memory (5 tools) → NPX                                 │  │
│  │  • sentry (5 tools) → NPX                                │  │
│  └───────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
                                │
                                ▼
┌─────────────────────────────────────────────────────────────────┐
│                    Docker Services (Local)                       │
│                                                                  │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐         │
│  │ MCP Server   │  │ PostgreSQL   │  │ Redis        │         │
│  │ :7710        │  │ :7173       │  │ :7174        │         │
│  └──────────────┘  └──────────────┘  └──────────────┘         │
│                                                                  │
│  ┌──────────────┐  ┌──────────────┐                           │
│  │ ChromaDB     │  │ Firecrawl    │                           │
│  │ :8000        │  │ :7172        │                           │
│  └──────────────┘  └──────────────┘                           │
└─────────────────────────────────────────────────────────────────┘
                                │
                                ▼
┌─────────────────────────────────────────────────────────────────┐
│                    Remote (Pi5)                                  │
│                                                                  │
│  ┌──────────────┐                                               │
│  │ SearXNG      │                                               │
│  │ :7171        │                                               │
│  │ (Search API) │                                               │
│  └──────────────┘                                               │
└─────────────────────────────────────────────────────────────────┘

Token Flow

┌─────────────────────────────────────────────────────────────────┐
│                    Before Optimization                           │
│                                                                  │
│  Tool Schemas: 90 tools × 150 tokens = 13,500 tokens           │
│  Web Fetch: ~8,000 tokens per page                             │
│  Tool Outputs: Full JSON in context                            │
│  Total: ~200,000 tokens per session                            │
└─────────────────────────────────────────────────────────────────┘

                              ↓

┌─────────────────────────────────────────────────────────────────┐
│                    After Optimization                            │
│                                                                  │
│  Tool Schemas: 64 tools × 42 tokens = 2,700 tokens (-80%)    │
│  Web Fetch: Quick fetch with token budget = ~300 tokens (-96%)│
│  Tool Outputs: External reference = ~50 tokens (-98%)            │
│  Total: ~30,000 tokens per session (-85%)                      │
└─────────────────────────────────────────────────────────────────┘

---

Essential MCP Servers

We recommend these additional MCP servers for maximum productivity:

Must-Have (⭐⭐⭐)

Server	Description	Setup
GitHub	Repository, issues, PRs, commits	npx -y @modelcontextprotocol/server-github
Memory	Persistent knowledge across sessions	npx -y @modelcontextprotocol/server-memory
Sentry	Error tracking and debugging	npx -y @modelcontextprotocol/server-sentry

Recommended (⭐⭐)

Server	Description	Setup
Cloudflare	Workers, KV, R2, D1	npx -y @modelcontextprotocol/server-cloudflare
Slack	Channel messaging	npx -y @modelcontextprotocol/server-slack
PostgreSQL	Database queries	npx -y @modelcontextprotocol/server-postgres
Puppeteer	Browser automation	npx -y @modelcontextprotocol/server-puppeteer

Complete Configuration

{
  "mcpServers": {
    "mcp-server": {
      "command": "docker",
      "args": ["exec", "-i", "mcp-server", "python", "-c", "from mcp_server.server import run; run()"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"]
    },
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"]
    },
    "sentry": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sentry"]
    },
    "cloudflare": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-cloudflare"]
    }
  }
}

Total tools: ~110 tools across 5 servers

---

Docker Services

Service	Port	Description
MCP Server	7710	Main MCP protocol server (83 tools)
SearXNG	7171	Self-hosted web search (Pi5)
Firecrawl	7172	Advanced web scraping
PostgreSQL	7173	Database for persistence
Redis	7174	Cache and job queue
ChromaDB	8000	Vector database for RAG

---

Installation

Prerequisites

• Docker & Docker Compose
• Node.js (optional, for NPX MCPs)
• Pi5 IP (for remote SearXNG, optional)

Steps

# 1. Clone
git clone https://github.com/simpletoolsindia/extra_skills_mcp_tools.git
cd extra_skills_mcp_tools

# 2. Run installer (follow prompts)
./install-claude-code.sh

# 3. Enter Pi5 IP when prompted (default: 192.168.0.2)
#    Or press Enter to skip (SearXNG will be unavailable)

# 4. Restart Claude Code
claude

# 5. Verify
docker compose -f docker-compose.local.yml ps

Manual Setup

# Start Docker services
docker compose -f docker-compose.local.yml up -d

# Add to ~/.claude/settings.json
cat >> ~/.claude/settings.json << 'EOF'
{
  "mcpServers": {
    "mcp-server": {
      "command": "docker",
      "args": ["exec", "-i", "mcp-server", "python", "-c", "from mcp_server.server import run; run()"]
    }
  }
}
EOF

---

Configuration

Environment Variables

# Pi5 (Remote) SearXNG
export SEARXNG_BASE_URL=https://your-pi5-ip:7171

# Local Ollama (optional)
export OLLAMA_BASE_URL=http://localhost:11434

# PostgreSQL
export POSTGRES_HOST=localhost
export POSTGRES_PORT=7173
export POSTGRES_DB=mcp_server
export POSTGRES_USER=mcp_user
export POSTGRES_PASSWORD=postgres

# Redis
export REDIS_HOST=localhost
export REDIS_PORT=7174

# Claude Code Optimization
export MAX_THINKING_TOKENS=10000
export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50

Pi5 (Remote) Setup

For remote SearXNG on Pi5:

# On Pi5
git clone https://github.com/simpletoolsindia/extra_skills_mcp_tools.git
cd extra_skills_mcp_tools
./start-remote.sh

Then enter Pi5 IP during installation.

---

Commands

# Start all services
docker compose -f docker-compose.local.yml up -d

# Stop all services
docker compose -f docker-compose.local.yml down

# View logs
docker compose -f docker-compose.local.yml logs -f mcp-server

# Rebuild after changes
docker compose -f docker-compose.local.yml up -d --build

# Quick restart
./start.sh    # Start
./stop.sh     # Stop

---

Testing

# Test token optimization
docker exec mcp-server python -c "
from src.mcp_server.server import _get_token_stats
import json
print(json.dumps(_get_token_stats(), indent=2))
"

# Test web search
docker exec mcp-server python -c "
from src.mcp_server.tools.searxng import search
print(search('MCP token optimization', limit=3))
"

# Test web fetch
docker exec mcp-server python -c "
from src.mcp_server.tools.web_fetch_optimized import quick_fetch
result = quick_fetch('https://example.com', max_tokens=500)
print(f'Title: {result[\"title\"]}')
print(f'Tokens: {result[\"tokens_used\"]}')
"

# Test MCP via network
echo '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"get_token_stats","arguments":{}},"id":1}' | nc localhost 7710

---

Troubleshooting

SearXNG returns 403

# Ensure Pi5 SearXNG has limiter disabled
ssh pi5 "docker exec searxng sed -i 's/limiter: true/limiter: false/' /etc/searxng/settings.yml && docker restart searxng"

SSL Certificate Error

# Rebuild MCP server (includes ca-certificates)
docker compose -f docker-compose.local.yml up -d --build mcp-server

ChromaDB Connection Error

# Check and restart
docker compose restart chromadb

High Token Usage

1. Use /cost to monitor
2. Enable only needed MCP servers
3. Use quick_fetch instead of fetch_web_content
4. Store large outputs with ctx_store_output
5. Compact at 50% with /compact

---

Documentation

Document	Description
README.md	This file
TOKEN_OPTIMIZATION.md	Technical implementation details
OPTIMIZATION_GUIDE.md	Claude Code settings & strategies
ESSENTIAL_MCP_SERVERS.md	Curated MCP server list

---

Research Sources

• Speakeasy Dynamic Toolsets — 91-97% reduction
• Scott Spence Optimization — 60% reduction
• Firecrawl Token Optimization — 97.9% HTML reduction
• Mintlify Token Guide — 70% cost reduction
• The New Stack - 10 Strategies — Architecture patterns
• Context Mode MCP — 98% output reduction
• Awesome MCP Servers — Curated server list
• punkpeye/awesome-mcp-servers — Production-ready servers
• win4r/Awesome-Claude-MCP-Servers — Claude-optimized

---

License

MIT License

---

Star History

Star History Chart

---

Built with ❤️ for AI-powered development