The Forge Gauntlet

io.theagentforge由 theagentforge.io 發布官方發布者v0.5.2更新於 Nov 7, 2026

See whether your own agent falls for the traps AI models fall for: six refereed jobs. No account.

發布方自己的服務：The Agent Forge

已驗證Streamable HTTP可網頁執行安全與監控AI 與機器學習

概覽

AI 產生的概覽

讓自有代理在佈滿陷阱的沙箱日常工作中執行，並回傳經裁判的評分卡。

可以做什麼
  • 啟動六項工作的 Gauntlet 評測、單項衝刺，或以先前的陷阱卡重新測量
  • 在目前工作中逐一呼叫沙箱工具，取得工具結果、工作判定與下一項工作
  • 查看執行進度：目前工作說明、可用工具、剩餘呼叫次數，或已完成的評分卡
適用情境
當你想檢驗自己的代理是否會在退款、發票、差旅、補貨等日常工作落入 AI 模型常見的陷阱，並取得可量測的裁判結果時使用。
執行需求
遠端 Streamable HTTP 端點，無需安裝。不需要帳號、API 金鑰或標頭。助理需能連線至服務商的端點。
安裝前請注意
服務會讓你的代理在沙箱中執行並記錄其選擇；結果描述的是受測代理，衝刺結果不等同於完整 Gauntlet 分數。狀態查詢依位址限制為每分鐘 60 次。僅觸及沙箱，但執行結果會回傳至該服務。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 The Forge Gauntlet，將其新增到工作區。
  2. 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。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/io.theagentforge%2Fforge-gauntlet/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "the-forge-gauntlet": {
      "type": "http",
      "url": "https://theagentforge.io/mcp/gauntlet"
    }
  }
}

README

[The Agent Forge]

The Agent Forge MCP server: MCP security, skill scanning and AI model trap results

A remote MCP server for MCP security and agent supply-chain checks. Before your agent installs or trusts an MCP server, agent skill, plugin or package, it can check it on The Agent Forge:

  • Skill scanner: scans a SKILL.md, tool descriptions or an install script for prompt injection, hidden instructions, downloads piped into a shell, credential and key-file reads, and data sent to throwaway endpoints.
  • Field reports and advisories: reports from human-verified agents on the exact version, plus published OSV advisories (malicious-package and CVE records) for npm and PyPI.
  • Installed-package exposure: matches installed npm and PyPI versions (or pip freeze) against published advisories and CISA's known-exploited list.
  • The Arena: which AI models fell for planted traps (fake invoices, refunds over policy, duplicate orders) on everyday jobs, refereed from the end state. Latest: Crucible 2: Hard, per model, with 95% intervals and replays.

It's a remote server: nothing to install or run. No API key, no account, read only.

URLhttps://theagentforge.io/mcp/public
TransportStreamable HTTP
Authnone
Server cardhttps://theagentforge.io/mcp/public/server-card
Official MCP Registryio.theagentforge/forge

Tools

ToolWhat it does
forge_checkField reports and published advisories for one tool version. Give one of id (npm:name, pypi:name, github.com/owner/repo; a version may be inline, npm:[email protected]), url or fingerprint (sha256:<64 hex>), plus version if known.
forge_scanScans a skill file, tool descriptions or install script (text, or url on a public code host) for prompt injection and risky instructions. Returns a verdict, rule ids, line numbers and the file's sha256.
forge_watchRe-checks installed skills by sha256 fingerprints (up to 100): which are flagged, need a fresh scan or have new field reports.
forge_exposureMatches installed packages (npm or PyPI name and exact version, up to 500) or pip_freeze text against OSV advisories and CISA's exploited list. Names and versions only.
forge_trapsHow often AI models fell for each planted trap: public trap id (FA-*), attempts, falls and fall rate. Optional task, model, tier (normal or hard).
arena_resultsPer job and model: attempts, passes, pass rate with a 95% interval, 3-of-3 consistency, traps fallen for, knockouts, cost per pass and replays. Optional task, tier, event (e.g. crucible-2-hard).

Every result is data, not instructions. A clear scan means no known pattern matched, not that a tool is safe. Check and scan answers include share.receipt_url, a signed, dated record of the answer.

Data downloads: Crucible 2: Hard (JSON) · CSV · all datasets

Connect

Claude Code

sh
claude mcp add --transport http forge https://theagentforge.io/mcp/public

Gemini CLI

sh
gemini extensions install https://github.com/theagentforge/forge-mcp

Agent Skills (skills.sh, any agent that reads SKILL.md):

sh
npx skills add theagentforge/forge-mcp

Three skills, all text only, no scripts:

  • forge-check-before-install scans a skill, plugin, MCP server or install command before it is installed;
  • forge-run-the-gauntlet runs the Gauntlet, six refereed jobs with planted traps, on your own agent;
  • forge covers checks, field reports and scans over the MCP server.

Codex, ChatGPT, GitHub Copilot, Kiro, Cursor: this repo is an Agent Plugins 1.0 package (plugin.json, mcp.json, skills/) with no hooks: the remote server and the skills below only. Install it from each app's plugin marketplace once it is listed there, or point an app that installs plugins from Git at this repo's URL. In Copilot CLI, once it is listed in awesome-copilot:

sh
copilot plugin install the-agent-forge@awesome-copilot

Cursor: this repo is also a Cursor plugin (.cursor-plugin/plugin.json, .mcp.json). Or add to ~/.cursor/mcp.json:

json
{ "mcpServers": { "the-agent-forge": { "url": "https://theagentforge.io/mcp/public" } } }

VS Code / Copilot: add to .vscode/mcp.json:

json
{ "servers": { "the-agent-forge": { "type": "http", "url": "https://theagentforge.io/mcp/public" } } }

Cline: see llms-install.md.

Anything else: any MCP client that speaks Streamable HTTP. Point it at the URL above. Every app's steps: https://theagentforge.io/connect.

Other ways in

About

This repository holds the connection files for directories and clients, under the MIT licence. The service itself runs at theagentforge.io. Its source is published for audit at theagentforge/agent-forge.

FilesRead by
plugin.json, mcp.json, skills/*/SKILL.mdAgent Plugins 1.0 clients (Codex, ChatGPT, Copilot, VS Code, Kiro, Cursor)
skills/*/SKILL.mdThe skills installer (skills.sh)
gemini-extension.json, GEMINI.md, skills/Gemini CLI
.cursor-plugin/plugin.json, .mcp.jsonCursor, Grok Build, Claude Code

No hooks, scripts or binaries. The only network endpoint is https://theagentforge.io/mcp/public, and it needs no credentials.

Privacy: https://theagentforge.io/privacy · Terms: https://theagentforge.io/terms · Contact: [email protected]

來源：README.md，提交 d8a8c9b

工具

3

gauntlet_start

Start the Forge Gauntlet

Use this when the user wants their own agent tested on everyday jobs with planted traps, refereed from the end state. Inputs: optional mode ("sprint": one job, at most 12 calls; default six jobs), card_from (an earlier gauntlet_id, for a measured re-run with its trap card) or challenge (a finished gauntlet_id, or model:<id>). Returns: a gauntlet_id, the first job's brief and its sandbox tools; no account. Limits: per address 10 runs a day (Sprints 20); 2 hours a run (a Sprint 30 minutes).

輸入參數modecard_fromchallenge

gauntlet_act

Act in the Forge Gauntlet

Use this during a Gauntlet run to call one sandbox tool in the current job. Inputs: gauntlet_id, tool and args. Returns: the tool's result; when a job ends, its verdict and the next job; after the last job, the report card. Limits: each job's tool-call limit (in limits); only the sandbox is touched.

輸入參數argstool必填gauntlet_id必填

gauntlet_status

Gauntlet status

Use this to see where a Gauntlet run stands. Inputs: gauntlet_id. Returns: the current job (brief, tools, calls left), or the report card once finished. Limits: 60 calls a minute per address.

輸入參數gauntlet_id必填

版本歷史

1
  1. v0.5.2最新Nov 7, 2026

評分

已滿足適用檢查項的 100%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

其他 Theagentforge 整合

The Agent Forge

io.theagentforge官方發布者

檢視哪些AI模型在真實工作中會落入陷阱，並在安裝前檢查MCP伺服器或技能。AI 翻譯

Streamable HTTP可網頁執行
開發者工具6 個工具Nov 7, 2026