3%/ day
daily_loss_limitNo new buys once today’s loss reaches 3% of your start-of-day equity.
Every order checked first. Loss limits, size caps and revenge-trade blocks.
I
How it works
An ordinary “guardrail” MCP server only advises. Your agent can still call the broker’s order tool directly. Anthropic Trade is a proxy: the agent connects only to Anthropic Trade, and Anthropic Trade holds the broker connection. Blocking is enforced, not suggested.
1 · Your agent
Plans a trade and calls place_order, the same as with any broker MCP.
2 · Anthropic Trade
Loss limit, size cap, revenge window, buy count, loss streak. Any failure stops the order and says why.
3 · Broker
Only orders that pass reach the broker. Every verdict goes into the journal.
Rules only stop orders that add risk. Cutting a loss or taking profit always goes through.
There is no tool for it. Only you change them, in the dashboard, and they apply from the next order.
The reply carries the numbers: today’s loss, the position’s share of equity, minutes left on a cooldown, the max size allowed.
II
Rules
The defaults follow common retail risk rules. Change any number in the dashboard, and hit one button to pause every buy.
3%/ day
daily_loss_limitNo new buys once today’s loss reaches 3% of your start-of-day equity.
5%max
position_sizeNo single stock above 5% of equity (broad funds like SPY or VTI: 30%). A blocked order is told the largest size it may buy instead.
15min
revenge_tradeNo buying a stock back within 15 minutes of selling it at a loss.
10buys / day
max_buys_per_dayCaps overtrading when the agent gets busy.
3losses → 60 min
loss_streak_cooldownThree losing exits in a row pause all new buys for an hour.
1click
kill_switchPause all buys from the dashboard, instantly, until you turn it back on.
Always
reduce_onlySells that reduce a position always pass. Anthropic Trade never traps you in a trade.
III
See it work
A real run on the simulated $100k account (npm run demo).
IV
Dashboard
A local dashboard at localhost:5900/app shows every order the agent tried, what got through and why the rest were stopped.
One button pauses all buys. Selling stays open so you can always get out.
Turn on dashboard approval and every order that passes the rules waits for your click before it reaches the broker.
Today’s loss budget, position sizes and the full journal, refreshed every second and a half.
V
Install
Needs Node 20 or newer. Nothing to clone: npm fetches Anthropic Trade when your AI app starts it.
claude mcp add anthropic-trade -- npx -y anthropic-trade
// mcpServers in the config file
"anthropic-trade": {
"command": "npx",
"args": ["-y", "anthropic-trade"]
}
npx anthropic-trade dashboard
# then open localhost:5900/app
Tools:
guard_statusget_accountget_quotecheck_orderplace_orderget_journal
VI
Modes
Default
A simulated $100,000 account that starts from real closing prices for AAPL, MSFT, NVDA, TSLA, AMZN, GOOGL, META, AMD, PLTR, COIN, HOOD and the funds SPY, QQQ, VTI, BND and GLD. No broker account needed.
Try asking your agent:Buy 500 HOOD, thenWhat’s my max size?
Experimental
Run npm run login once to connect your Robinhood account through Robinhood’s own sign-in (OAuth). Anthropic Trade then relays Robinhood’s hosted trading MCP: order tools are checked first, everything else passes through unchanged.
Any Robinhood tool Anthropic Trade doesn’t recognise as a read, a cancel or an order is refused, so nothing slips past. P&L is estimated from the orders Anthropic Trade forwarded. Not yet tested on a live account.
Start in simulation, tune your rules, then connect your broker.
Add to Claude