Agentrisk MCP server

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 checktraderisk, 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

Shortened. The full README is on GitHub.