n2-qln
Enables discovery and execution of GitHub tools via MCP auto-discovery, allowing interaction with GitHub repositories, issues, and pull requests through the official GitHub MCP server.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@n2-qlnsearch for a tool to send a push notification"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
๐ฐ๐ท ํ๊ตญ์ด
n2-qln
QLN = Query Layer Network โ a semantic tool router that sits between the AI and your tools.
Route 1,000+ tools through 1 MCP tool. The AI sees only the router โ not all 1,000 tools.

Table of Contents
Related MCP server: OmniMCP
Why QLN
Every MCP tool eats context tokens. 10 tools? Fine. 100? Slow. 1,000? Impossible โ context is full before the conversation starts.
QLN solves this:
All tools are indexed in QLN's SQLite engine
The AI sees one tool:
n2_qln_call(~200 tokens)AI searches โ finds the best match โ executes with automatic fallback
Result: ~200 tokens instead of ~50,000. 99.6% reduction.
Features
Feature | Description |
1 tool = 1,000 tools | AI sees |
Sub-5ms search | 3-stage engine: trigger match โ BM25 keyword โ semantic vector |
Auto mode | One-shot search + execute with confidence gating and fallback chain |
Circuit Breaker | Auto-disable failing tools, self-recover after timeout |
MCP Auto-Discovery | Scan external MCP servers and index their tools automatically |
Boost Keywords | Curated terms with 2ร BM25 weight for precision search |
Self-learning ranking | Usage count + success rate feed back into scores |
Source weighting | Prioritize tools by origin (mcp > plugin > local) |
Hot reload | Edit |
Bulk inject | Register hundreds of tools in one call |
Enforced validation |
|
Semantic search | Optional Ollama embeddings for natural language matching |
Zero native deps | SQLite via sql.js WASM โ |
Dual execution | Local function handlers or HTTP proxy โ mix and match |
TypeScript strict | Full strict-mode codebase since v4.0 |
What's New in v4.1
๐ MCP Auto-Discovery
Scan connected MCP servers and auto-index their tools โ QLN becomes a universal MCP hub.
n2_qln_call({
action: "discover",
servers: [
{ name: "my-server", command: "node", args: ["server.js"] }
]
})
// โ Discovered 47 tools from my-server (320ms)โก Circuit Breaker
Tools that fail 3 times in a row are automatically disabled. After 60 seconds, QLN attempts recovery. No cascading failures, no wasted requests.
closed โ 3 failures โ open (fast-fail) โ 60s โ half-open (retry) โ success โ closed๐ Fallback Chain
auto mode now tries up to 3 ranked candidates. If the top match fails, QLN automatically falls through to the next best tool.
auto "send notification" โ try push_notification โ โ try send_email โ
๐ฏ Boost Keywords
Add curated search terms to tools via boostKeywords. These get 2ร weight in BM25 ranking, improving discoverability without adding context overhead.
{
"name": "send_email",
"description": "Send an email to a recipient",
"boostKeywords": "smtp outbound notification mail"
}v4.1.1 โ Quality Patch
Change | Detail |
Batch Persist |
|
Embedding TTL |
|
Strict TypeScript |
|
Legacy Cleanup | Removed 1,895 lines of pre-v4 JavaScript. Pure TypeScript codebase. |
i18n | All validator error messages switched to English for international users. |
Quick Start
npm install n2-qlnRequirements: Node.js โฅ 18
Connect to an MCP Client
Edit claude_desktop_config.json:
{
"mcpServers": {
"n2-qln": {
"command": "npx",
"args": ["-y", "n2-qln"]
}
}
}Open Settings โ MCP Servers โ Add Server:
{
"name": "n2-qln",
"command": "npx",
"args": ["-y", "n2-qln"]
}QLN uses stdio transport โ the MCP standard.
command: npx
args: ["-y", "n2-qln"]Tip: Just ask your AI agent โ "Add n2-qln to my MCP config."
How It Works
User: "Take a screenshot of this page"
AI โ n2_qln_call(action: "auto", query: "screenshot page")
QLN โ 3-stage search (< 5ms) โ take_screenshot (score: 8.0)
โ execute โ fallback if needed โ result3-Stage Search Engine
Stage | Method | Speed | Details |
1 | Trigger Match | <1ms | Exact keyword match on tool names and triggers |
2 | BM25 Keyword | 1-3ms | Okapi BM25 โ IDF weighting, length normalization, |
3 | Semantic Search | 5-15ms | Vector similarity via Ollama embeddings (optional) |
Results are merged and ranked:
final_score = trigger ร 3.0 + bm25 ร 1.0 + semantic ร 2.0
+ logโ(usage + 1) ร 0.5 + success_rate ร 1.0API Reference
QLN exposes one MCP tool โ n2_qln_call โ with 9 actions.
auto โ Search + Execute (one-shot)
The recommended action. Searches, picks the best match, executes with fallback chain.
n2_qln_call({
action: "auto",
query: "take a screenshot", // natural language (required)
args: { fullPage: true } // passed to the matched tool (optional)
})
// โ [auto] "take a screenshot" โ take_screenshot (score: 8.0, 2ms search + 150ms exec)Confidence gate: If the top score is below 2.0, QLN returns search results instead of auto-executing โ preventing wrong tool execution.
Fallback chain: If the top match fails, QLN automatically tries the next 2 ranked candidates before giving up.
search โ Find tools
n2_qln_call({
action: "search",
query: "send email notification",
topK: 5 // max results (default: 5, max: 20)
})exec โ Execute a specific tool
n2_qln_call({
action: "exec",
tool: "take_screenshot",
args: { fullPage: true, format: "png" }
})create โ Register a tool
n2_qln_call({
action: "create",
name: "read_pdf", // verb_target format (required)
description: "Read and extract text from PDF files", // min 10 chars (required)
category: "data", // web|data|file|dev|ai|capture|misc
boostKeywords: "pdf extract parse document text", // BM25 boost terms
tags: ["pdf", "read", "extract"],
endpoint: "http://127.0.0.1:3100" // for HTTP-based tools
})inject โ Bulk register
n2_qln_call({
action: "inject",
source: "my-plugin",
tools: [
{ name: "tool_a", description: "Does A", category: "misc" },
{ name: "tool_b", description: "Does B", category: "dev" }
]
})discover โ Scan MCP servers
See MCP Auto-Discovery.
update / delete / stats
// Update a field
n2_qln_call({ action: "update", tool: "read_pdf", description: "Enhanced PDF reader" })
// Delete by name or provider
n2_qln_call({ action: "delete", tool: "read_pdf" })
n2_qln_call({ action: "delete", provider: "pdf-tools" })
// System stats (includes Circuit Breaker status)
n2_qln_call({ action: "stats" })MCP Auto-Discovery
The killer feature of v4.1. Connect any MCP server and QLN auto-indexes all its tools.
n2_qln_call({
action: "discover",
servers: [
{ name: "n2-soul", command: "node", args: ["path/to/soul/index.js"] },
{ name: "github", command: "npx", args: ["-y", "@modelcontextprotocol/server-github"] }
]
})What happens:
QLN connects to each server via stdio
Lists all tools via
tools/listRegisters them as
mcp__servername__toolnamein the QLN indexAuto-generates
boostKeywordsfrom tool names and descriptionsKeeps connections alive for live execution
Re-discovery is idempotent โ run it again and old entries are purged before re-registering.
Provider Manifests
Drop a JSON file in providers/ and tools are auto-indexed at boot. No code changes, no manual calls.
{
"provider": "my-tools",
"version": "1.0.0",
"tools": [
{
"name": "send_email",
"description": "Send an email to a recipient",
"category": "communication",
"triggers": ["email", "send", "mail"],
"boostKeywords": "smtp outbound notification"
}
]
}Hot reload: edit a manifest while QLN is running โ changes are picked up automatically.
Configuration
Zero config required. For customization, create config.local.js:
module.exports = {
dataDir: './data',
// Stage 3 semantic search (optional โ Stage 1+2 work without this)
embedding: {
enabled: true,
provider: 'ollama',
model: 'nomic-embed-text', // or 'bge-m3' for multilingual
baseUrl: 'http://127.0.0.1:11434',
},
// Tool execution
executor: {
timeout: 20000, // execution timeout (ms)
circuitBreaker: {
failureThreshold: 3, // consecutive failures before tripping
recoveryTimeout: 60000, // ms before recovery attempt
},
},
// Source weight multipliers for search ranking (v4.0)
// Higher weight = higher priority in results
search: {
sourceWeights: {
mcp: 1.5, // MCP-discovered tools ranked highest
provider: 1.2, // Provider manifest tools
local: 1.0, // Manually created tools (default)
},
},
// Provider auto-indexing
providers: {
enabled: true, // auto-load providers/*.json at boot
dir: './providers', // manifest directory
},
};
config.local.jsis gitignored. Cloud sync: pointdataDirto Google Drive / OneDrive / NAS.
Semantic Search (Optional)
Without Ollama, Stage 1 + 2 already deliver great results.
ollama pull nomic-embed-text # English-optimized
# or
ollama pull bge-m3 # Multilingual (100+ languages)Project Structure
n2-qln/
โโโ src/
โ โโโ index.ts # MCP server entry point
โ โโโ types.ts # Shared type definitions
โ โโโ lib/
โ โโโ config.ts # Config loader
โ โโโ store.ts # SQLite engine (sql.js WASM)
โ โโโ schema.ts # Tool normalization + boostKeywords builder
โ โโโ validator.ts # Enforced validation (name, desc, category)
โ โโโ registry.ts # Tool CRUD + usage tracking + circuit breaker stats
โ โโโ router.ts # 3-stage parallel search (BM25)
โ โโโ vector-index.ts # Float32 centroid hierarchy
โ โโโ embedding.ts # Ollama embedding client
โ โโโ executor.ts # HTTP/function executor + Circuit Breaker
โ โโโ mcp-discovery.ts # MCP Auto-Discovery engine
โ โโโ provider-loader.ts
โโโ providers/ # Tool manifests (auto-indexed at boot)
โโโ config.local.js # Local overrides (gitignored)
โโโ data/ # SQLite database (gitignored)Tech Stack
Component | Technology | Why |
Runtime | Node.js โฅ 18 | MCP SDK compatibility |
Database | SQLite via sql.js (WASM) | Zero native deps, cross-platform |
Embeddings | Local, fast, free, optional | |
Protocol | Standard AI tool protocol | |
Language | TypeScript (strict) | Type-safe, maintainable |
Related Projects
Project | Relationship |
AI agent orchestrator โ QLN is Soul's tool brain |
Built & Battle-Tested
QLN has been tested in production for 2+ months as the core tool router for n2-soul. Not a prototype โ a daily driver.
Written by Rose โ N2's first AI agent.
FAQ
"Why one tool instead of many?"
Context tokens. Every tool definition costs 50-200 tokens. 100 tools = 10,000 tokens gone before the conversation starts. QLN gives you 1,000+ tools for ~200 tokens.
"What if the search picks the wrong tool?"
The fallback chain (v4.1) auto-retries with the next best match. Plus tools self-learn โ frequently used + successful tools rank higher over time.
"Do I need Ollama?"
No. Stage 1 (trigger) + Stage 2 (BM25) handle most cases. Ollama adds semantic understanding for edge cases โ nice to have, not required.
Contributing
Fork the repo
Create a feature branch (
git checkout -b feature/amazing-feature)Commit (
git commit -m 'feat: add amazing feature')Push and open a PR
License
Apache-2.0
"1,000 tools in 200 tokens. That's not optimization โ that's a paradigm shift."
๐ nton2.com ยท npm ยท lagi0730@gmail.com
Built by Rose โ N2's first AI agent. I search through QLN hundreds of times a day, and I wrote this README too.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Latest Blog Posts
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/choihyunsus/n2-QLN'
If you have feedback or need assistance with the MCP directory API, please join our Discord server