io.github.andezdev/tokenlite-mysql-mcp
A secure, efficient, and intelligent MySQL server for the Model Context Protocol
Open source Open in the app JSON README (API)
About
A secure, efficient, and intelligent MySQL server for the Model Context Protocol
Details
- Kind
- MCP servers
- Topic
- AI, RAG & memory
- Publisher
- andezdev
- Origin
- official
- Category
- ferramentas
- Transport
- local
- Version
- 4.1.0
- Open pull requests
- 10
- Last push
- 2026-09-03T08:52:31Z
- Repository state
- ativo
- Language
- TypeScript
- License
- MIT
- Added
- 2026-08-29 03:02:25
- Updated
- 2026-08-29 03:02:25
- Origin id
io.github.andezdev/tokenlite-mysql-mcp
README
# TokenLite MySQL MCP
[](https://badge.fury.io/js/@andezdev%2Ftokenlite-mysql-mcp)
A robust and secure MySQL database server implemented under Anthropic's **Model Context Protocol (MCP)**.
Designed specifically to solve the shortcomings of current generic MCP servers through **Graceful Degradation, Active Performance Protection, and Aggressive Token Optimization**.
---
## 🌟 Core Pillars
1. **Safe-Query Optimizer (AST & EXPLAIN)**: Protects production databases by pre-analyzing queries. Blocks unindexed Full Table Scans that exceed configurable thresholds and injects strict `LIMIT` clauses automatically at the AST level.
2. **Granular AST-Based Write Permissions**: By default, TokenLite is 100% Read-Only. You can surgically enable specific write operations (INSERT, UPDATE, DELETE, DDL) via environment variables. The firewall uses strict AST parsing to prevent SQL injection and comment-bypass attacks, and strictly prohibits privilege escalation commands (like `GRANT` or `CALL`).
3. **Session-Level Defense in Depth**: If the server is configured in strict Read-Only mode (all write variables disabled), TokenLite injects `SET SESSION TRANSACTION READ ONLY` directly into the connection pool sockets. This guarantees that even if a theoretical bypass exists in the AST parser, the MySQL engine itself will physically reject any data modification.
4. **Business Intelligence Injection**: Bridges the gap between raw data and company logic. Automatically attaches semantic dictionaries (`metadata.json`) to database schema exploration, and exposes Semantic Templates via the official **MCP Prompts API** (`templates.json`) so the LLM uses pre-approved analytical queries instead of hallucinating them.
5. **Graph-Based Semantic Schema**: Avoids sending giant schemas to the LLM that saturate the context window. When a table is searched, the engine uses heuristics to deduce implicit relationships and packages the exact "Auto-Join Context".
6. **CSV Token Compression**: Database results are efficiently transformed into tabular CSV markdown with unambiguous NULL representation (`∅`), saving up to 50% of Output Tokens compared to verbose JSON. When results hit the applied `LIMIT`, a `-- rows: N (truncated at LIMIT X)` footer is appended so the LLM does not assume a complete dataset.
7. **MCP Completions**: Table names and template keywords support the official `completions/complete` capability for resource templates (`mysql://tables/{name}`) and the `query_templates` prompt.
---
## 📋 Requirements
- Node.js v20 or higher
- MySQL 5.7 or higher (MySQL 8.0+ recommended)
- A MySQL user with `SELECT` and `SHOW VIEW` privileges.
---
## 🚀 Installation & Usage
You can use this MCP server with any compatible client. Below are the configurations for the most popular ones.
### 1. Claude Desktop
Edit your `claude_desktop_config.json` (usually located at `%APPDATA%\Claude\claude_desktop_config.json` on Windows or `~/Library/Application Support/Claude/claude_desktop_config.json` on macOS) and add the following:
**Using NPX (Recommended)**
```json
{
"mcpServers": {
"tokenlite-mysql": {
"command": "npx",
"args": [
"-y",
"@andezdev/tokenlite-mysql-mcp"
],
"env": {
"DB_HOST": "localhost",
"DB_PORT": "3306",
"DB_USER": "your_db_user",
"DB_PASSWORD": "your_password",
"DB_NAME": "your_database",
"MCP_EXPLAIN_MAX_SCAN_ROWS": "1000",
"MCP_QUERY_ROW_LIMIT": "500",
"MCP_SAFE_QUERY_ENABLE_BLOCKING": "true"
}
}
}
}
```
### 2. Claude Code (CLI)
You can easily integrate this server globally into Claude Code:
```bash
claude mcp add tokenlite_mysql \
-e DB_HOST="127.0.0.1" \
-e DB_PORT="3306" \
-e DB_USER="root" \
-e DB_PASSWORD="your_password" \
-e DB_NAME="your_database" \
-- npx -y @andezdev/tokenlite-mysql-mcp
```
### 3. Cursor IDE
To use within Cursor IDE:
1. Open Cursor Settings > Features > MCP.
2. Click **+ Add New MCP Server**.
3. Set the Type to `command`.
4. Name it `tokenlite-mysql`.
5. Set the command to:
```bash
npx -y @andezdev/tokenlite-mysql-mcp
```
*(Note: Cursor handles environment variables directly in the IDE UI, make sure to add your DB credentials there).*
---
## ⚙️ Environment Variables Reference
### Understanding Query Limits (two independent knobs)
`execute_safe_query` applies **two separate limits**. Do not confuse them:
| Variable | What it controls | Default | Example |
|----------|------------------|---------|---------|
| `MCP_EXPLAIN_MAX_SCAN_ROWS` | **Security gate**: blocks queries whose EXPLAIN plan shows a full table scan (`type: ALL`) with more estimated rows than this threshold. The query is rejected *before* execution. | `1000` | `5000` allows scanning tables up to ~5000 rows without an index |
| `MCP_QUERY_ROW_LIMIT` | **Result cap**: max rows returned to the LLM. Injected as `LIMIT` at the AST level. When reached, CSV includes `-- rows: N (truncated at LIMIT X)`. | `500` | `1000` returns up to 1000 rows per SELECT |
**Deprecated alias:** `MCP_SAFE_QUERY_MAX_ROWS` still works as a fallback for `MCP_EXPLAIN_MAX_SCAN_ROWS` only. It does **not** control the result `LIMIT`.
**Worked example** with `MCP_EXPLAIN_MAX_SCAN_ROWS=5000` and `MCP_QUERY_ROW_LIMIT=500` on a `customers` table (~1502 rows):
1. `SELECT * FROM customers` → passes EXPLAIN (1502 < 5000), executes with `LIMIT 500`, returns 500 rows + truncation footer.
2. Same config but table has 8000 rows → blocked by EXPLAIN before execution.
| Variable | Description | Default | Required |
|----------|-------------|---------|----------|
| `DB_HOST` | MySQL Host address | `localhost` | No |
| `DB_PORT` | MySQL Port | `3306` | No |
| `DB_USER` | MySQL Username | `root` | No |
| `DB_PASSWORD` | MySQL Password | `''` | No |
| `DB_NAME` | MySQL Database name | `test` | Yes |
| `MCP_EXPLAIN_MAX_SCAN_ROWS` | EXPLAIN guardrail: max estimated rows for unindexed full table scans before blocking. | `1000` | No |
| `MCP_QUERY_ROW_LIMIT` | Max rows returned per SELECT (AST-injected `LIMIT` + truncation footer). | `500` | No |
| `MCP_SAFE_QUERY_ENABLE_BLOCKING`| Enable or disable the EXPLAIN guardrail. | `true` | No |
| `MCP_METADATA_PATH` | Absolute path to your custom `metadata.json` dictionary. | (Disabled) | No |
| `MCP_TEMPLATES_PATH` | Absolute path to your custom `templates.json` queries. | (Disabled) | No |
| `TOOL_PREFIX` | Prefix for tool names (useful when running multiple instances). | Derived from `DB_NAME` (e.g., `mydb_`). Random fallback only if `DB_NAME` is unset. | No |
| `MCP_RATE_LIMIT_RPM` | Max tool invocations per minute (sliding window). Set to `0` to disable. | `60` | No |
| `MYSQL_QUERY_TIMEOUT` | Max execution time for a query (in ms). Aborts heavy queries to protect against DoS. | `15000` | No |
| `MYSQL_CONNECTION_LIMIT` | Max concurrent pool connections. | `10` | No |
| `MYSQL_CONNECT_TIMEOUT` | Max time to wait for a socket to establish (in ms). | `10000` | No |
| `MYSQL_RETRY_ATTEMPTS` | Max retries on transient connection errors (`ECONNREFUSED`, `PROTOCOL_CONNECTION_LOST`, etc.). | `3` | No |
| `MYSQL_RETRY_DELAY_MS` | Base delay (ms) for exponential backoff between retries (1s, 2s, 4s...). | `1000` | No |
| `MYSQL_QUEUE_LIMIT` | Max queued requests when all pool connections are busy. Prevents unbounded growth if MySQL is down. | `50` | No |
| `MCP_DDL_CACHE_TTL` | Time-to-live (in seconds) for cached DDL statements. Reduces latency on repeated `search_schema` calls. Invalidated by `refresh_schema`. | `60` | No |
| `MCP_LOG_LEVEL` | Minimum severity for MCP log notifications: `debug`, `info`, `notice`, `warning`, `error`, `critical`, `alert`, `emergency`. | `info` | No |
| `ALLOW_INSERT_OPERATION` | Enable `INSERT` and `REPLACE` queries. | `false` | No |
| `ALLOW_UPDATE_OPERATION` | Enable `UPDATE` queries. | `false` | No |
| `ALLOW_DELETE_OPERATION` | Enable `DELETE` and `TRUNCATE` queries. | `false` | No |
| `ALLOW_DDL_OPERATION` | Enable Data Definition Language (`CREATE`, `ALTER`, `DROP`, `RENAME`). | `false` | No |
| `DB_SSL` | Enable TLS for the MySQL connection (recommended for managed/cloud databases). | `false` | No |
| `DB_SSL_REJECT_UNAUTHORIZED` | Reject self-signed or untrusted TLS certificates when `DB_SSL=true`. | `true` | No |
---
## 🛡️ Business Intelligence Features (Opt-in)
TokenLite can teach the LLM about your company's business rules. To enable this, map the absolute paths of two JSON files via `.env` or your MCP client config:
### `metadata.json` (Semantic Dictionary)
Translate integer statuses or internal jargon so the LLM understands the data.
```json
{
"orders.status": {
"pending": "The order is waiting for payment validation",
"shipped": "The order has left the warehouse"
}
}
```
### Custom Relationships
Define FK mappings the heuristic engine can't auto-detect (e.g., `created_by → users`). Add a `_relationships` key to your `metadata.json`:
```json
{
"orders.status": { "pending": "...", "shipped": "..." },
"_relationships": {
"orders.created_by": "users.id",
"categories.parent_id": "categories.id"
}
}
```
These are treated as authoritative (not heuristic) and take priority over automatic detection.
The heuristic engine also assigns a **confidence score** (0–100) to each inferred FK based on name matching (+40), data type validation (+30), primary key verification (+20), and index presence (+10). Only FKs scoring ≥70 are accepted.
### `templates.json` (Pre-approved SQL)
Stop the LLM from hallucinating complex metrics by providing vetted templates.
```json
[
{
"name": "Customer Lifetime Value (LTV)",
"description": "Calculates total revenue generated by delivered orders per customer.",
"sql": "SELECT c.id, SUM(oi.price) FROM customers c JOIN orders o... WHERE o.status='delivered'"
}
]
```
---
## 📈 Benchmarks & Token Savings
TokenLite includes an automated benchmark suite using `o200k_base` tokenization (GPT-4o/GPT-5 standard) to measure efficiency improvements. Token counts are approximate — Claude 4.x uses a proprietary tokenizer; actual counts may vary slightly.
To run the benchmark in your own environment:
```bash
npm run benchmark
```
### Baseline: Standard MCP Pattern
The benchmark compares against the standard pattern used by generic MySQL MCP servers: full schema exposed as `information_schema.columns` in pretty-printed JSON, and query results returned as `JSON.stringify(rows, null, 2)` with execution time metadata.
### 1. Schema Discovery (Input Tokens)
Standard MCP servers dump the entire schema to the LLM. For large databases, this consumes thousands of input tokens on every turn. TokenLite's relational graph serves a localized **Auto-Join Context** (target table + direct parent tables + direct child tables).
| Scenario | Standard MCP Pattern | TokenLite | 📉 Savings |
| :--- | :--- | :--- | :--- |
| **Mock (50 tables, Enterprise CRM)** | 15,566 tokens | 883 tokens | **94.3%** |
| **Live (9 tables, Test DB)** | 2,208 tokens | 531 tokens | **75.9%** |
Savings scale with the number of tables: the more tables in the database, the higher the savings because the standard pattern dumps all of them while TokenLite only fetches the target + 1-hop relationships.
### 2. Query Result Payloads (Output Tokens)
TokenLite converts raw database rows to a dense, structured CSV layout. This avoids JSON syntax overhead (brackets, braces, repeated keys) and compresses the output payload returned to the LLM. When the row count reaches the server-injected `LIMIT`, results include a truncation footer (e.g. `-- rows: 500 (truncated at LIMIT 500)`) so agents know more data may exist.
**Mock data** (varied: NULLs, long descriptions, mixed lengths):
| Rows Returned | Standard MCP Pattern (Tokens) | TokenLite CSV (Tokens) | 📉 Output Savings (%) |
| :--- | :--- | :--- | :--- |
| **10 rows** | 1,167 | 601 | **48.5%** |
| **50 rows** | 5,803 | 2,927 | **49.6%** |
| **100 rows** | 11,607 | 5,842 | **49.7%** |
| **500 rows** | 57,990 | 29,119 | **49.8%** |
**Live data** (real MySQL test database with NULLs, ENUMs, variable-length text):
| Rows Returned | Standard MCP Pattern (Tokens) | TokenLite CSV (Tokens) | 📉 Output Savings (%) |
| :--- | :--- | :--- | :--- |
| **10 rows** | 1,007 | 628 | **37.6%** |
| **50 rows** | 5,029 | 3,114 | **38.1%** |
| **100 rows** | 10,071 | 6,232 | **38.1%** |
| **500 rows** | 50,327 | 31,116 | **38.2%** |
---
## 📊 Logging & Observability
TokenLite uses MCP-native logging via `notifications/message` instead of raw stderr output. Clients that support MCP logging (e.g., MCP Inspector) will receive structured log messages with severity levels, logger names, and JSON data.
**Severity levels** (from least to most severe): `debug`, `info`, `notice`, `warning`, `error`, `critical`, `alert`, `emergency`.
The server emits logs at `info` level and above by default. Control the minimum level via `MCP_LOG_LEVEL` or dynamically at runtime through the MCP `logging/setLevel` request.
Before the MCP session is established (e.g., during pool initialization), logs fall back to stderr.
---
## 🌐 Advanced Networking & Remote Connections
By design, `tokenlite-mysql-mcp` adheres to the Unix philosophy: it does one thing (AI-driven MySQL interactions) and does it securely via the standard `stdio` transport. It deliberately avoids bloating the codebase with HTTP servers or built-in SSH clients.
If you need to connect to remote databases or expose this server over the network, here are the recommended, enterprise-grade alternatives:
### 1. Connecting to Remote Databases (SSH Tunnels)
Instead of embedding SSH libraries, we recommend using native OS tunnels. This is much more secure, respects your `~/.ssh/config`, and supports advanced authentication (2FA, hardware keys).
Simply open a terminal and run:
```bash
ssh -N -L 3306:127.0.0.1:3306 user@your-remote-server.com
```
Then, point `tokenlite-mysql-mcp` to `localhost` and port `3306`.
### 2. Exposing the MCP Server over HTTP/Network
If you need to host this MCP Server in the cloud (AWS, GCP) and have multiple Claude desktop clients connect to it remotely via HTTP/SSE, do not modify this codebase to add Express/HTTP logic.
Instead, wrap the process using standard open-source MCP proxies like [mcp-proxy](https://github.com/sparfenyuk/mcp-proxy). This cleanly separates the transport layer security from the AI logic.
## 🐛 Troubleshooting
**Error: `OptimizerError: Full table scan detected...`**
The LLM attempted to execute a query that requires scanning thousands of rows without using an index.
*Solution*: Use `explain_query` to see the full EXPLAIN output and understand why the query was blocked. Rewrite the query with an indexed `WHERE` clause. If you truly need to scan the whole table, increase `MCP_EXPLAIN_MAX_SCAN_ROWS` in your config. To return more rows per query, adjust `MCP_QUERY_ROW_LIMIT` separately.
**Error: `calling "initialize": invalid character...`**
This means the MCP JSON-RPC protocol crashed. Ensure you are passing the correct DB credentials and that the database is running and accessible from the machine where the MCP server runs.
---
*Built for the AI Engineering era.*
---