trading-agent-safety

SecProve — Trading Agent Safety (Claude plugin)

Safety-first Claude skills for AI agentic trading. If you're connecting Claude to Robinhood's agentic trading via MCP, this plugin makes sure your agent trades inside limits you chose — caps, an approval gate, a kill switch, and prompt-injection defense — applied automatically, before the first order.

Built by SecProve, where cybersecurity skill is measured. The guidance mirrors the Trading Agent Safety guide.

---

⚠️ Read this first

• This plugin helps you bound an autonomous trading agent. It does not pick trades, does not give financial advice, and cannot remove the risk of agentic trading — which includes the possible loss of your entire funded balance.
• AI-driven strategies can perform poorly, move faster than you can react, and be hard to stop in real time. Fund only a separate agentic account, with money you can afford to lose entirely.
• The strategy skills are educational playbooks, not recommendations. You are responsible for everything your agent does.
• Read what can go wrong before you trade.

This is software for risk reduction, not a money-making system.

---

What's inside

One plugin, trading-agent-safety, with six skills. The guardrails skill is the point; the strategy skills all refuse to run without it.

Skill	Invocation	What it does
agent-guardrails	automatic	Applies per-trade / daily / concentration caps, an approval gate, a circuit breaker, a kill switch, and a prompt-injection rule whenever Claude operates the Robinhood Trading MCP. The safety net.
mean-reversion	/trading-agent-safety:mean-reversion	Bounded mean-reversion playbook
buy-the-dip-dca	/trading-agent-safety:buy-the-dip-dca	Bounded DCA / buy-the-dip playbook
rebalancing	/trading-agent-safety:rebalancing	Bounded portfolio-rebalancing playbook
momentum	/trading-agent-safety:momentum	Bounded momentum playbook
sector-rotation	/trading-agent-safety:sector-rotation	Bounded sector-rotation playbook

The strategy skills are user-invoked only (disable-model-invocation: true) so Claude never auto-starts trading, and each one requires agent-guardrails active.

Install

/plugin marketplace add secprove/trading-agent-safety
/plugin install trading-agent-safety@secprove

Use it safely (new users)

1. Install the plugin (above).
2. Connect the Robinhood Trading MCP — see how to connect Claude.
3. Tell Claude your funding amount and risk tier (conservative / balanced / aggressive). agent-guardrails sizes the caps and enforces them.
4. Optional: run a bounded strategy, e.g. /trading-agent-safety:momentum.

Full walkthrough: https://secprove.com/trading-agent-safety/agent-safety-skill

How the guardrails behave

Hard caps that can't be exceeded · approval required on large orders · circuit breaker halts a runaway · STOP kill switch · refuses rule-changing input it reads (prompt-injection defense) · equities-only during the Robinhood beta.

License

MIT © SecProve. Contributions welcome.

Could you spot a prompt-injection attempt aimed at your own agent? Test your security instincts at secprove.com.