MCP Examples

Purpose

Provide real-world MCP patterns from production servers: code search (lci), browser integration (agnt), process management, and knowledge bases.

When to Use

• Need concrete examples of patterns
• Want to see actual implementations
• Designing similar functionality
• Learning from working systems

Code Search Pattern (lci)

Architecture

• Pattern: Hub-and-Spoke + Progressive Discovery
• Tools: 8+ tools
• Token System: result_id, symbol_id

Key Tools

search - Hub tool

{
  "input": {"pattern": "string", "filter": "optional"},
  "output": {
    "results": [
      {"id": "r1", "name": "User.authenticate", "preview": "...", "conf": 0.95}
    ],
    "has_more": true,
    "total": 127
  }
}

get_definition - Spoke tool

{
  "input": {"id": "r1"},
  "output": {
    "symbol_id": "s1",
    "name": "User.authenticate",
    "signature": "...",
    "source": "...",
    "location": {"file": "user.ts", "line": 42}
  }
}

Token efficiency: ID reference saves ~80% tokens vs. repeating full code

Progressive Detail Example

Query: "authenticate"

High match (0.95): Full details (200 tokens)
  - Name, signature, docs, preview, location

Medium match (0.70): Summary (50 tokens)
  - Name, type, file

Low match (0.40): Minimal (10 tokens)
  - Name, ID only

Browser Proxy Pattern (agnt)

Architecture

• Pattern: CRUD + Aggregation
• Tools: 10+ tools
• Token Systems: proxy_id, session_id, request_id

Key Tools

proxy_start - Create

{
  "input": {"target_url": "http://localhost:3000"},
  "output": {
    "proxy_id": "dev",
    "listen_addr": "http://localhost:12345",
    "status": "running"
  }
}

currentpage - Aggregation

{
  "input": {"proxy_id": "dev"},
  "output": {
    "session_id": "page-1",
    "url": "http://localhost:3000",
    "errors_count": 3,          // Not full error objects
    "interactions_count": 127,   // Not every interaction
    "mutations_count": 45,       // Not every mutation
    "performance": {...}
  },
  "detail_access": "Use detail=['errors'] for full data"
}

Key pattern: Counts in overview, full data on request

Hierarchical IDs

proxy_id (dev)
  ↓
session_id (page-1)
  ↓
request_id (req_a1b2)

Each level provides more specificity.

Process Management Pattern

Architecture

• Pattern: CRUD + Lazy Loading
• Tools: 8+ tools
• Token System: process_id

Progressive Status

Level 1 - Count

{
  "active": 5,
  "stopped": 2
}

Level 2 - List

{
  "processes": [
    {"id": "p1", "name": "dev-server", "status": "running"},
    {"id": "p2", "name": "test", "status": "running"}
  ]
}

Level 3 - Status

{
  "id": "p1",
  "status": "running",
  "uptime": "2h15m",
  "memory": "245MB",
  "preview": "Server listening :3000"
}

Level 4 - Full

{
  /* ...all Level 3... */,
  "full_output": "... complete logs ...",
  "env": {...},
  "metrics": {...}
}

Knowledge Base Pattern

Architecture

• Pattern: Discovery-Detail
• Tools: Search, topics, articles
• Token System: article_id, topic_id

Layered Access

list_topics()
  → ["auth", "deploy", "monitor"]

get_topic_summary("auth")
  → {articles: 12, updated: "2024-01"}

search_articles("OAuth")
  → [{id: "a1", title: "...", preview: "..."}]

get_article("a1")
  → {title, content, related: [...]}

Common Patterns Across Examples

1. ID Reference System

All use IDs to avoid repeating data:

• lci: result_id → symbol_id
• agnt: proxy_id → session_id → request_id
• process: process_id
• kb: topic_id → article_id

Savings: 70-90% token reduction

2. Progressive Detail

All vary detail by context:

• lci: By confidence (0.95 = full, 0.40 = minimal)
• agnt: By request (counts vs. full arrays)
• process: By depth (count → list → status → full)
• kb: By layer (topics → summary → full article)

3. Automation Flags

All include standard flags:

{
  "has_more": boolean,
  "total": integer,
  "returned": integer,
  "complete": boolean
}

4. Accept Extra Parameters

All accept unknown params with warnings:

const {known, params, ...extra} = input
if (extra) warnings.push(`Unknown: ${Object.keys(extra)}`)

Anti-Patterns Seen and Fixed

❌ Repeating Data

Before (wasteful):

// Tool 1
{"results": [{"name": "...", "code": "... 200 lines ..."}]}

// Tool 2 needs same data
// User copies entire result

After (efficient):

// Tool 1
{"results": [{"id": "r1", "name": "...", "preview": "10 lines"}]}

// Tool 2
input: {"id": "r1"}  // Reference only

❌ No Progressive Detail

Before:

{
  "results": [
    {"name": "...", "full": "... 500 tokens ..."},
    {"name": "...", "full": "... 500 tokens ..."},
    {"name": "...", "full": "... 500 tokens ..."}
  ]
}

After:

{
  "results": [
    {"id": "a1", "conf": 0.95, "full": "..."},  // Only high confidence
    {"id": "b2", "conf": 0.70, "summary": "..."},
    {"id": "c3", "conf": 0.40}  // Just ID
  ]
}

❌ Flat Structure

Before (15+ tools, no organization):

search_users, search_posts, get_user, get_post, ...

After (grouped):

Query Tools: search
Lookup Tools: get_user, get_post
Management: create_user, update_user

Real-World Token Savings

lci code_search Tool

Without IDs:

• Average result: 250 tokens (full code)
• 10 results: 2,500 tokens

With IDs:

• Average preview: 50 tokens
• 10 results: 500 tokens
• Savings: 80%

agnt currentpage Tool

Without aggregation:

• Full errors array: 400 tokens
• Full interactions: 600 tokens
• Full mutations: 300 tokens
• Total: 1,300 tokens

With aggregation:

• Error count: 10 tokens
• Interaction count: 10 tokens
• Mutation count: 10 tokens
• Total: 30 tokens (97% savings)
• Use detail parameter for full arrays when needed

Quick Reference

Proven patterns:

1. Hub-and-Spoke - lci (search → details)
2. CRUD - agnt (lifecycle management)
3. Aggregation - agnt currentpage (counts not arrays)
4. Lazy Loading - process status (overview → full)
5. Discovery-Detail - kb (topics → articles)

Key lessons:

• IDs save 70-90% tokens
• Progressive detail by relevance/confidence
• Counts in overview, arrays on request
• Accept extra params with warnings
• Automation flags for AI agents

Study these real-world examples when designing similar functionality.