One question, one answer

Is there an MCP server for crypto trading?

Answered from the server’s own tool table: what the MCP endpoint is, which tools need a key, and what an agent connected to it can and cannot do.

The full Bot API referenceWhere to test a trading botThe keyless market-data APITry the calls live

What you actually connect to

The Model Context Protocol lets an assistant call tools on a server instead of guessing from training data. MarginPad runs one at https://marginpad.io/mcp over Streamable HTTP, so there is nothing to install, nothing to host and no local process to keep alive. In a client that accepts remote MCP servers the whole configuration is one object:

{"mcpServers":{"marginpad":{"url":"https://marginpad.io/mcp"}}}

A machine-readable descriptor sits at /.well-known/mcp.json if your tooling prefers to discover it rather than be told.

Two halves: reading the market, and trading a paper account

The market-data half needs no key at all. An assistant can ask for the price of BTC, the last hour of candles, current funding across venues, open interest, the long/short split, today’s liquidations, a scored screener or the economic calendar, and get a measured number back. Those calls are keyless because the data is ours to give: the liquidations come from our own collector subscribed to nine exchanges’ public websockets, not resold from a vendor.

The trading half operates a paper futures account: open a leveraged position at the live price, close part of it, move a stop, rest a limit or stop order, read the balance, the open book and the trading report. Those tools take a free API key in an X-API-Key header, minted from the Bot API page with an email and nothing else.

The split matters for trust. An agent can be given the reading half with no credential at all, and the trading half only when you actually want it placing orders.

What an agent connected to this cannot do

It cannot deposit, withdraw, or touch an exchange account, because there is no route to one: MarginPad holds no custody, takes no deposits and asks for no exchange API keys. Positions settle on our servers in simulated dollars against real live prices.

That is the whole reason to hand an agent this endpoint rather than a live exchange key. An agent that misreads a signal, loops on an order, or sizes a position ten times too large costs you a log line here. The same mistake on a funded account costs the account. Forward-test the agent where the consequences are removed and the prices are not.

Why the answers are worth more than a scrape

An assistant answering from a scraped page is quoting whatever was cached the last time a crawler passed. These tools return the figure at the moment of asking, with a timestamp, from the same feed our own charts and paper terminal run on.

Two habits are worth building into any agent that uses this: call get_economic_calendar for event dates rather than answering from memory, since a model’s idea of when CPI lands is frequently a year stale, and call get_markets before sizing a trade, because leverage caps and fee rates differ per symbol and asset class.

If you would rather not use MCP

Everything the MCP server does is also a plain HTTP call. The same account is reachable over REST, the OpenAPI 3.1 spec is at /openapi.json, positions can be streamed over WebSocket instead of polled, and signed webhooks push fills and liquidations to a URL you control. Official clients with no dependencies install with pip install marginpad or npm install marginpad.

The free plan carries all of it: 120 requests a minute per key, three keys, fifty open positions, the WebSocket stream, replay of a past day and every keyless market-data endpoint. Paid plans raise ceilings and add webhooks and AI market reads - they do not unlock the product.

Questions

Related: the full Bot API reference · where to test a trading bot · the keyless market-data API · try the calls live