# sdk [Health: Active]

**Category:** 💰 Finance & Fintech  
**Repository:** https://github.com/marginpad/sdk  
**GitHub Stars:** 0  
**npm Downloads (last month):** 17054  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/sdk-2

## Description
Free crypto-futures paper trading and live market data for AI agents. 29 tools; no key for data.

## Claude Desktop Quick Installation
No confirmed install details yet, so no config is published. Follow the setup steps in the repository above.

## Documentation

## What sdk does

The sdk MCP server connects AI agents to MarginPad’s crypto-futures paper-trading API and its live market-data endpoints. The service is designed for testing bots and strategies with live prices while keeping execution inside a simulated venue. Market-data access does not require a key; trading operations use a MarginPad API key.

The server exposes 29 tools according to the listing. Its supported workflows include opening, managing, and closing positions; reading account state; configuring books; inspecting prices; and receiving trading updates. The paper venue models fees, funding, liquidation conditions, order types, and other execution behavior, but it does not send orders to a real exchange.

This fits the Finance & Fintech category for developers building agent-driven trading experiments, strategy evaluations, or paper accounts that need more than static historical data.

## How it works

An agent sends trading and data requests through the MCP server to MarginPad. A key can be associated with a separate book, allowing one strategy to keep an isolated journal and result set. The API supports idempotent client order IDs, so a retry can return the existing position rather than create another one.

Simulation behavior can be adjusted across account, fill, cost, and liquidation settings. Examples include starting balance, leverage limits, margin enforcement, slippage, latency, partial fills, rejection rates, taker fees, funding, and maintenance-margin rules. Presets such as realistic, Bybit-style, and brutal provide starting configurations. Values outside the documented ranges are rejected rather than silently changed.

Replay mode runs a key through a prior day using candle data. Trading calls act on the replay journal, while stops, targets, and liquidations are evaluated between calls. Equity curves, maximum drawdown, replay cursors, candles, and closing summaries are available through the SDK’s related operations.

## Setup and configuration

The sdk MCP server is documented for Claude, ChatGPT, and Cursor. The repository includes an MCP example for Claude Desktop or Cursor, and the project links to MarginPad’s trading API guide and OpenAPI reference for service details.

The underlying Python package requires Python 3.8 or newer and has no dependencies. The optional `stream` extra adds WebSocket support. The JavaScript package requires Node 18 or newer and has zero dependencies. Python and JavaScript clients receive the API key in the `MarginPad` constructor; the README does not specify environment variables for MCP configuration.

Free accounts allow three keys, one book, 50 open positions, and 120 requests per minute. Paid API plans increase these limits and add webhook capacity and longer history retention. Webhooks are not available on the Free plan.

## Tools and capabilities

The sdk MCP server supports capabilities including:

- Reading keyless crypto prices and market data
- Opening and managing long or short paper positions
- Setting stop-loss, take-profit, and trailing-stop behavior
- Reading accounts, books, equity curves, and drawdown
- Configuring simulation and execution assumptions
- Starting, inspecting, and stopping historical replays
- Streaming price and position-change events

## Limitations and notes

Paper results are not live-exchange results. Market impact is modeled from configured assumptions instead of measured against an order book, and stop, target, or liquidation fills do not include crossing costs. A strategy that performs well in this environment still needs a cautious live test.

Rate limits and plan limits apply. Responses include rate-limit headers, and throttled clients are expected to wait and retry once. Trade history is retained for 30 days on the Free plan and 90 days on paid plans. The project is MIT licensed and describes the simulator as educational rather than financial advice.

_Full upstream README: https://allmcps.com/mcp/sdk-2/readme_