AgentRisk

Risk guardrails for AI trading agents. Your agent proposes; your policy decides.

AgentRisk sits in front of trade execution: it analyzes portfolio risk, checks each proposed trade against a policy you write, and manages that policy. The core is deterministic (same inputs, same verdict), transparent (policies are plain YAML you can read and edit), and fail-closed (invalid input or a missing policy blocks, never a silent pass).

AgentRisk never recommends trades and never executes them. A PASS means a trade did not break the rules you wrote, not that it is safe or profitable. See DISCLAIMER.md.

Quickstart

The fastest path for a Claude agent is the Claude Code plugin. It installs the Agent Skill and registers the MCP server in one step (via uvx, so there is nothing else to install):

/plugin marketplace add trycoin-ai/agentrisk
/plugin install agentrisk@agentrisk

For any other MCP client, register the server yourself:

{
  "mcpServers": {
    "agentrisk": {
      "command": "uvx",
      "args": ["--from", "agentrisk[mcp]", "agentrisk-mcp"]
    }
  }
}

Then ask questions in plain English and the agent translates them into three tool calls:

The plugin path needs uv for uvx. See examples/claude_desktop.md for a two-minute Claude Desktop setup.

Agent Skill

For Claude agents, AgentRisk ships an optional Agent Skill that encodes the discipline the guardrail depends on: before any order reaches a broker, classify the trade, call check_trade_risk, respect the verdict, and record the result, in that order. The plugin above installs it with the server. To install just the skill:

• Claude Code or Claude Desktop: copy skills/agentrisk/ into ~/.claude/skills/.
• claude.ai: upload skills/agentrisk/SKILL.md as a skill.

The enforcement contract

AgentRisk returns advice. It cannot physically stop an order, so your integration must gate execution on the verdict:

result = check_trade_risk(portfolio, trade)
if not result.proceed:
    refuse(result.summary)              # BLOCK: never call the broker
elif result.acknowledgements_required:
    confirm_with_user(result)           # WARN: surface warnings first
else:
    execute(trade)                      # PASS

If you call the broker regardless of the verdict, you have a logger, not a guardrail. See the integration guide.

Using the library directly

The three tools are also a pure Python library, with no network calls and no LLM in the core:

pip install agentrisk

Every parameter, return field, and error case is in the tool reference.

What it checks

• Concentration caps on single names, sectors, themes, and asset classes, evaluated on the simulated post-trade portfolio.
• Asset-class rules (allow, warn, or block) for crypto, options, and margin.
• Order sanity: max order size, minimum cash floor, insufficient-funds detection.
• Restricted symbols and data quality (stale snapshots warn, invalid ones block).

The policy reference covers every field and the safety behaviors: fail closed, exits are never trapped, only breach-worsening trades block, and the one-time bypass.

Documentation

Roadmap

Next is deterministic stress scenarios, ETF look-through, and behavioral limits (v0.2), then options analytics, broker snapshot adapters, short positions, and multi-currency (v0.3). Trade recommendations, signals, execution, and telemetry are permanently out of scope. See the milestones.

Contributing

Classification-data corrections and new deterministic checks are especially welcome. See CONTRIBUTING.md, and keep the core pure: no network, no LLM calls, no hidden state.

License

MIT. See LICENSE.