Rein Agent Risk Scale

io.github.reinlayerv0.1.0更新於 Oct 7, 2026

Indicative Agent Risk Rating (A-E) self-check for AI agents that spend money. Not a credit rating.

已驗證Streamable HTTP可網頁執行AI & ML Security & Monitoring Finance

概覽

AI 產生的概覽

讓助理對可花錢的 AI 代理進行自報的 A-E 風險自檢，並解釋評級含義。

功能: 這個遠端 MCP 端點提供三個工具：self_check 依結構化回答為部署評分，get_scale 回傳 A-E 等級與各等級含義，explain_grade 說明某個等級需要滿足什麼。評分採用固定的公開規則而非語言模型，因此相同回答永遠得到相同等級。回應會列出最能解釋該等級的發現，以及可提升等級的做法。
適用情境: 當你希望助理評估或記錄可花錢代理周邊的控制措施時使用，例如問責人、支出上限、收款方限額、停止開關、人工升級、監控與事件歷史。這是自報的指示性檢查，不是信用評級，也不是投資建議。
執行需求: 遠端 streamable HTTP MCP 端點；未宣告登入、API 金鑰或環境變數。任何 MCP 用戶端都可連線該端點網址。README 中的範例腳本使用 Python，代理範例還需要 OpenAI API 金鑰。

安裝前請注意

等級為自報且未經核實，因此在控制措施被核實前一直是 E（未核實）。回答與等級以假名化紀錄儲存，只有設定 publish_opt_in 為 true 時才會公布識別碼。切勿傳送機密或個人資料。

安裝

在 SourceWeft 中

開啟儀表板中的 Rein Agent Risk Scale，將其新增到工作區。
為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。遠端服務在工作區中設定後即可從網頁執行環境執行。

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "rein-agent-risk-scale": {
      "type": "http",
      "url": "https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/mcp"
    }
  }
}

README

Rein Agent Risk Scale: MCP examples

Copy-paste examples that connect an AI agent to the Rein Agent Risk Scale self-check over MCP. The self-check gives an agent that can spend money an indicative Agent Risk Rating on an A-E scale, from the controls around it: who is accountable, spend caps and where they are enforced, payee limits, a stop switch and who holds it, human escalation, monitoring, history and incidents.

MCP endpoint (streamable HTTP, no sign-in):

The endpoint has three tools:

self_check scores a deployment from structured answers;
get_scale returns the A-E scale and what each grade means;
explain_grade says what a grade requires.

Scoring is fixed, published rules with no language model, so the same answers always give the same grade.

Examples

Each example is one file of under 40 lines.

file	framework	install (tested versions)
`examples/plain_mcp.py`	the `mcp` Python SDK, no LLM	`pip install "mcp==2.3.0"`
`examples/openai_agents.py`	OpenAI Agents SDK	`pip install "openai-agents==0.23.1"`
`examples/langgraph_agent.py`	LangChain / LangGraph with the MCP adapters	`pip install "langchain==1.4.3" "langchain-openai==1.6.7" "langchain-mcp-adapters==0.3.2"`
`examples/crewai_agent.py`	CrewAI	`pip install "crewai==1.15.23" "crewai-tools[mcp]==1.15.23"`

The three agent examples use OpenAI by default (export OPENAI_API_KEY=...). Pass any model your framework accepts to main() to use another.

Tested: every example was run against the live endpoint on 7 October 2026 at the versions above.
Watched weekly: a check runs them against the latest framework releases, because releases rename things. mcp 2.x renamed its HTTP client, and it now yields two streams instead of three.

Use it from any MCP client

json

{  "mcpServers": {    "rein-agent-risk-scale": {      "type": "http",      "url": "https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/mcp"    }  }}

{  "mcpServers": {    "rein-agent-risk-scale": {      "type": "http",      "url": "https://rein-scale.mangowave-ad191d9c.eastus.azurecontainerapps.io/mcp"    }  }}

There are docs written for agents at /skill.md and /llms.txt.

What a self-check is, and is not

Indicative. The grade comes from your own answers. It is self-declared and unverified, so the rated grade stays E (unverified) until the controls are verified.
Explained. The response lists the findings that most explain the grade and the steps that would raise it.
Private by default. Answers and grades are stored under a pseudonymised record. A handle is published only if you set publish_opt_in: true. Never send secrets or personal data.
Not a credit rating, not investment advice, and not an offer of credit.

License

MIT. See LICENSE.

來源：README.md，提交 536abb4

工具

工具後設資料尚未被收錄。

版本歷史

v0.1.0最新Oct 7, 2026