Skip to main content
The platform ships a built-in MCP server, so AI agents and assistants (Claude, ChatGPT, Cursor, custom agents) can work with Hyperprop directly — no integration code required.
The key you connect with decides what the agent can do: Partner keys (hpk_...) don’t work with the connector.

Claude and ChatGPT

Add a custom connector: paste the endpoint URL and leave the OAuth Client ID/Secret fields empty — the connector registers itself automatically. A Hyperprop sign-in page opens where you paste your personal key or, for firms, an organization API key once. (Under the hood: OAuth 2.1 with PKCE and dynamic client registration.)
  • Personal key: the app gets its own connection with that key’s access; the app never receives the key itself. Each connected app is listed in Settings → API keys → Connected apps with when it was connected and last used. Disconnect it there at any time; revoking the key disconnects every app using it.
  • Organization key: access follows the key and ends when you revoke it.

Cursor / Claude Code

Add to mcp.json:
Traders put their personal key (hpp_...) in place of the organization key.

For traders: your own agent

Create a personal API key in Settings → API keys → Personal Keys and connect with it. The connector reaches all your accounts and orders, on crypto perpetuals and futures, like personal keys. Market data isn’t included: futures positions and accounts don’t show live prices or unrealized P&L (they carry "market_data_withheld": true).
Want an agent that can look but never trade? Connect a Read only key. The trading tools are then not offered at all.
Read — get_connection_info, list_accounts, list_positions, list_open_orders, list_order_history, list_fills, get_order_status, list_instruments, get_mark_prices (crypto perpetuals), get_perp_margin, get_account_lockout, get_risk_settings, get_journal_trades, get_journal_orders, get_journal_stats Trade (Read and write keys only): Every trading tool returns a command_id; get_order_status shows whether it was accepted, rejected or executed. Safeguards on agent trading:
  • Your approval first. Every trading tool requires confirm: true, and the agent is instructed to show you the exact action and wait for your approval before sending it.
  • Account-wide actions need an explicit account. cancel_all_orders, close_all_positions and flatten_account always act on the one account you name, and their descriptions spell out what they cancel or close.
  • Sizes in the instrument’s unit (for example 0.01 BTC, or whole contracts for futures), checked against the instrument’s size step and its minimum and maximum order sizes.
  • Per-order value limit for every order placed or reversed through the connector. Crypto orders are valued at size × price; market orders and reversals at the current mark price, and when no current price is available they are refused with PRICE_UNAVAILABLE (limit orders still work). Futures orders are valued at limit or stop price × contract multiplier × contracts. Larger orders can be placed in the Hyperprop app; list_instruments returns the limit as maxOrderValue.
  • Futures orders must be limit or stop orders. Market data isn’t included, so futures market orders and reversals can’t be checked and are refused with LIMIT_ORDER_REQUIRED. Closing futures positions works as usual.
  • Trading limits per minute, on top of your account’s usual order limits (see the table below). Exits have their own allowance, so orders never use up your ability to cancel or close.
  • Retries are safe. The connector makes each trading action idempotent, so a retried call isn’t executed twice.
  • Recorded in your account history. Every trading action an agent takes is logged as made through an AI agent, including attempts refused by a safeguard such as the rate limit or the order value limit. A call without confirm: true is not sent and not recorded.
  • Your account’s trading rules and risk settings apply exactly as they do in the app.
Errors carry a code and a hint with the next step, for example LIMIT_ORDER_REQUIRED, READ_ONLY_KEY, ORDER_TOO_LARGE, PRICE_UNAVAILABLE, CONFIRMATION_REQUIRED or RATE_LIMITED.

Connector limits

Each connected app made with a key shares that key’s allowance. Over a trading limit, the tool returns RATE_LIMITED with the seconds to wait. Over a message limit, the request answers HTTP 429 with JSON-RPC error -32000 and a Retry-After header. A request can carry at most 10 messages (more answers 400), and one client address can send at most 600 messages a minute. Repeated attempts with a wrong personal key answer 429; wait a minute.

Guide and prompts

On connect, the server gives the agent short trading instructions. It also publishes a trading guide as the resource hyperprop://guide (workflow, units, order types, brackets and rules) and these prompts:
  • review_my_journal — reviews your recent trades and statistics without trading.
  • account_overview — summarizes balances, lockout status, positions and working orders.
  • plan_order — works out a valid order with take profit and stop loss for a contract and checks it against the limits, without placing it.

For firms: operate your organization

Twelve tools, split by the permission they need. Read — list_trading_plans, list_trading_accounts, get_trading_account, get_account_audit_log, list_traders, get_trader, list_purchases, get_purchase, get_billing_summary Write — create_trading_account, update_trading_account, record_payout Every tool call executes the corresponding REST endpoint with your key, so organization scoping, permissions, validation, audit logging, and webhooks apply exactly as documented on each endpoint. An organization key can send 120 messages a minute, and the batch and address limits above apply too.
Read tools work with any key; write tools need write/admin permission. Connect a read-only key if you want a strictly read-only agent.
create_trading_account accepts an optional idempotencyKey so agent retries can’t create duplicates, and record_payout implements the documented payout recipe (append to metadata.payouts, adjust currentBalance, audit-logged reason).