Last updated: July 2026

Introduction

The Model Context Protocol (MCP) is an open standard that lets AI assistants connect to external tools through a standardized interface. Instead of copy-pasting schema outputs or writing ad hoc scripts, an MCP-connected AI can query your database directly, discover tables, run a replication, or copy files between cloud storage systems — all from a chat prompt.

Sling CLI ships a built-in MCP server. One command, sling mcp, exposes Sling’s full data-movement capability to any MCP-compatible client: Claude Desktop, Claude Code, GitHub Copilot in VSCode, or any other host that supports the protocol.

This article covers what the Sling MCP server exposes, how to wire it up in each client, and what you can actually do with it once it is running.

What the Sling MCP Server Exposes

Running sling mcp starts a stdio-based MCP server with six tools:

Each tool accepts an action parameter and an input object. The AI assistant constructs these calls from your natural-language prompt and executes them through the MCP layer — you see the result, not the JSON plumbing.

The server also ships three guided prompts for API spec workflows:

• api_spec_create_spec — builds a full API spec from documentation
• api_spec_add_endpoint — adds one endpoint to an existing spec
• api_spec_debug_endpoint — diagnoses and fixes a broken endpoint

Installing Sling

If you have not installed Sling yet:

# macOS / Linux
curl -fsSL https://slingdata.io/install.sh | bash

# Windows
irm https://slingdata.io/install.ps1 | iex

# Python
pip install sling

Confirm the install:

sling --version

Then set up your connections. Sling reads from ~/.sling/env.yaml:

connections:
  MY_PG:
    type: postgres
    host: db.example.com
    user: analyst
    password: "{{ env.DB_PASSWORD }}"
    database: production

  MY_S3:
    type: s3
    bucket: my-data-bucket
    access_key_id: "{{ env.AWS_ACCESS_KEY_ID }}"
    secret_access_key: "{{ env.AWS_SECRET_ACCESS_KEY }}"
    region: us-east-1

Test that your connections work before wiring up MCP:

sling conns test MY_PG
sling conns test MY_S3

Wiring Sling into Your AI Client

All clients use the same MCP config shape: point at the sling binary with args: ["mcp"] and pass your CLI Pro token as an environment variable.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "sling": {
      "command": "sling",
      "args": ["mcp"],
      "env": {
        "SLING_CLI_TOKEN": "your-cli-pro-token"
      }
    }
  }
}

Restart Claude Desktop. The MCP indicator (🔌) will appear in the bottom-right corner of the chat input when the server is connected.

Claude Code

Add the server at user scope (available in all projects) or project scope (team-shared via .mcp.json):

# User scope — global, for personal use
claude mcp add sling --scope user --args "mcp" --env SLING_CLI_TOKEN=your-token

# Project scope — team-shared
# Create .mcp.json in project root:

{
  "servers": {
    "sling": {
      "type": "stdio",
      "command": "sling",
      "args": ["mcp"],
      "env": {
        "SLING_CLI_TOKEN": "${SLING_CLI_TOKEN}"
      }
    }
  }
}

With ${SLING_CLI_TOKEN} the token is read from the environment at runtime, so it never lands in the config file committed to source control.

Verify the connection:

claude mcp logs sling

VSCode with GitHub Copilot

Create .vscode/mcp.json in your project root:

{
  "servers": {
    "sling": {
      "type": "stdio",
      "command": "sling",
      "args": ["mcp"],
      "env": {
        "SLING_CLI_TOKEN": "your-cli-pro-token"
      }
    }
  }
}

Open GitHub Copilot Chat, select Agent mode from the dropdown, then click the Tools button. The Sling connection, database, file_system, replication, and pipeline tools will appear in the list.

What You Can Do: Concrete Examples

Once the MCP server is running, you interact with it through natural-language prompts. Here are representative use cases with the actual tool calls the AI assistant constructs underneath.

Query a Database

Your prompt:

Use sling connection MY_PG to show me the top 10 products by revenue this month from the sales table

What the assistant executes:

{
  "action": "query",
  "input": {
    "connection": "MY_PG",
    "query": "SELECT product_id, SUM(revenue) AS total_revenue FROM sales WHERE date >= date_trunc('month', current_date) GROUP BY product_id ORDER BY total_revenue DESC LIMIT 10"
  }
}

Discover What Tables Exist

Your prompt:

Using sling, show me all tables in MY_PG that start with "customer_"

What the assistant executes:

{
  "action": "discover",
  "input": {
    "connection": "MY_PG",
    "pattern": "*.customer_*"
  }
}

The assistant can also call get_schemata to fetch full column-level metadata for a table before writing a query — useful when you do not have the schema memorized.

Compare Two Tables

Your prompt:

Use sling to compare MY_PG.dbt_dev.core_transactions (dev) with MY_PG.finance.core_transactions (prod) — row counts, null counts, distinct counts

The assistant calls database.query twice and returns a diff summary. This replaces a manual psql session with a single sentence.

Copy Files Between Cloud Storage Systems

Your prompt:

Use sling to copy all CSV files from MY_S3 folder raw/2025/ to MY_AZURE processed/ folder

What the assistant executes:

{
  "action": "copy",
  "input": {
    "source_location": "MY_S3/raw/2025/*.csv",
    "target_location": "MY_AZURE/processed/",
    "recursive": true
  }
}

Run a Replication

Your prompt:

Use sling to run the replication defined in /data/pipelines/users_replication.yaml

What the assistant executes:

{
  "action": "run",
  "input": {
    "file_path": "/data/pipelines/users_replication.yaml"
  }
}

If the YAML has a syntax error, the assistant can also call replication.parse first to validate it before running.

Build an API Spec from Scratch

This is where the guided prompts shine. Tell the assistant:

Use the api_spec_create_spec prompt with spec_name="stripe_custom", connection_name="MY_STRIPE",
api_docs_url="https://docs.stripe.com/api", endpoint_names="customers,charges,subscriptions"

The MCP prompt fetches Sling’s spec documentation, browses the Stripe docs, creates the YAML file, creates the connection, and iterates until all three endpoints return data successfully. What would take 30-60 minutes of manual work runs in under 5 minutes.

A Note on the Demo Video

The Sling team recorded a full walkthrough of the MCP server in action. It covers the Claude Desktop setup, a live database query, and a replication run end-to-end:

Debugging and Logs

If a tool call fails or returns unexpected results, check the MCP logs before assuming a connection issue:

For trace-level output from Sling itself, add DEBUG: TRACE to the env block in your MCP config:

{
  "mcpServers": {
    "sling": {
      "command": "sling",
      "args": ["mcp"],
      "env": {
        "SLING_CLI_TOKEN": "your-token",
        "DEBUG": "TRACE"
      }
    }
  }
}

This logs every tool action and its parameters to stderr, which the MCP client captures in the log file above.

Requirements

• Sling CLI installed and on PATH
• A valid CLI Pro token (SLING_CLI_TOKEN) — the MCP server requires a token to authorize tool calls. Tokens are available at app.slingdata.io.
• At least one connection configured in ~/.sling/env.yaml
• An MCP-compatible client (Claude Desktop, Claude Code, VSCode + Copilot)

Further Reading

• Sling MCP documentation
• Sling CLI getting started
• Configuring connections
• Replication guide
• API spec guide

Related guides

The MCP tools drive the same replications and queries you’d run by hand. These walkthroughs show the underlying commands the assistant constructs for you:

• Extract data from databases into DuckLake for a multi-database extraction the replication tool can run
• Import S3 CSV files into PostgreSQL for the kind of file copy the file_system tool handles
• Replicate PostgreSQL to DuckDB for a database-to-database move
• Sling vs Fivetran for how the single-binary design compares to a managed connector

Frequently asked questions

What is the Sling MCP server?

It’s a built-in Model Context Protocol server that ships with the Sling CLI. Running sling mcp starts a stdio-based server that exposes Sling’s data-movement tools to any MCP-compatible client such as Claude Desktop, Claude Code, or VSCode with GitHub Copilot.

Which tools does sling mcp expose?

Six tools: connection, database, file_system, api_spec, replication, and pipeline. Each takes an action and an input object. The server also ships three guided prompts for building, extending, and debugging API specs.

Do I need a CLI Pro token to use the Sling MCP server?

Yes. The MCP server requires a valid SLING_CLI_TOKEN to authorize tool calls. Tokens are available at app.slingdata.io. Pass it in the env block of your MCP client config.

How do I add the Sling MCP server to Claude Code?

Run claude mcp add sling with user or project scope, pass mcp as the argument, and set SLING_CLI_TOKEN in the env. For team sharing, add a stdio server entry to a project .mcp.json instead.

How do I keep my CLI Pro token out of source control?

Reference it as an environment variable in the config, for example SLING_CLI_TOKEN: "${SLING_CLI_TOKEN}". The token is read from the environment at runtime, so the committed config file never contains the secret.

How do I debug a failing MCP tool call?

Check the client MCP logs first: claude mcp logs sling for Claude Code, the ~/Library/Logs/Claude/ folder for Claude Desktop, or the MCP: Show Logs command in VSCode. For trace-level Sling output, add DEBUG: TRACE to the env block of your MCP config.

Can the AI assistant run a full replication through MCP?

Yes. The replication tool parses, compiles, and runs replication YAML files. Point the assistant at a YAML path and it calls the run action; it can call parse first to validate the file before running it.