Grill

io.github.mtangoz由 GitHub 账号 mtangoz 发布v0.1.0更新于 Oct 24, 2026

A second opinion from an unrelated AI model before you commit to a decision.

概览

AI 生成的概览

把你描述的决策发给另一家公司提供的 AI 模型,返回一句简短结论、最有力的反驳和一个低成本验证方法。

可以做什么
  • 提交你原样陈述的决策,获得简短的第三方结论
  • 得到一行证伪提示,指出会改变判断的关键问题
  • 获得针对方案最尖锐的一点和一个低成本测试,总长约 50 词
  • 在结尾行之后可要求更多细节、备选方案或你自己的判断
  • 查看运行说明,列出作答的模型和耗时
适用场景
在敲定方案、调价、上线时间等判断性决定之前,想要一个来自与你所用助手不同公司的反向意见时使用。在 Claude Code 即将批准计划时也适用。
运行要求
以 MCP 包形式在 Claude Desktop 本地运行,或在 Cursor、VS Code、Codex、Gemini CLI 等应用中通过 npx 运行。需要 OpenRouter API 密钥,从 GRILL_API_KEY 读取(未设置时使用 OPENROUTER_API_KEY),并需要 OpenRouter 余额。需能访问 OpenRouter;托管方式还需注册账号。
安装前请注意
你确认的书面描述会离开本机,发送给 OpenRouter 和外部模型;其中若含密钥、令牌或密码会中止运行,邮箱、电话、银行卡与银行账号、证件号、标注的出生日期、街道地址、邮编和 IP 会被遮蔽,但姓名不会。质量检查也会看到遮蔽后的内容,除非跳过。模型调用费用由你支付给路由服务。匿名使用统计默认关闭;可能出现新闻提示行和可选套餐提示。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Grill,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

[图片] Grill: grill your AI before you send it to your boss

Hear the strongest case against your plan, then decide for yourself.

Setup guide: grillyour.ai No key needed: Anyone can sign up and connect Grill to the AI they use at https://grillyour.ai/mcp. See it first: four made-up decisions, grilled

Tell the assistant you think with, "grill this": Claude, ChatGPT, Copilot, DeepSeek, Gemini, Grok or Muse. It sends the decision as you said it. An outside judge, a model from a different company, comes back with a short note: one verdict line, then one short line naming the question that would change the call when the judge named one, then the sharpest point and one cheap test when the judge named one. The verdict, the point and the test stay within about 50 words. The extra line is at most about 25 words, and it never comes before the verdict. Say more, options, or your call when you want the rest. After that closing line, a two-line run note names the models that answered and how long it took, then one sentence on privacy. Never drop the run note. Never put it before the verdict. A thin line gets a quick take and one question, not a refusal. A record is written only after you answer, and a percent is optional. Grill does not store it on the installed tool.

Verdict: shaky. The plan assumes customers stay at the new price, and nothing in it tests that. Falsifier: show the new price to one in ten new signups for two weeks, and compare how many start paying.

What Grill promises

The verdict never becomes the decision. You decide. Your values and judgment calls are yours: each verdict lists what the judge checked for you separately from what is your call. The judge is a model from a different company. Grill uses OpenRouter's Auto Router, and it does not pin models. Only the write-up you approve leaves your machine. Every promise, and the code and test that enforce it: docs/PRINCIPLES.md.

Set up

Claude Desktop and Claude Code enforce that in code. The copy-and-paste routes check it and warn you. Who to paste into is in the table below.

Grill in chat

Anyone can sign up. The steps are at grillyour.ai/start. Use the same email each time, the email you signed up with, every time.

  • Claude.ai, verified working. On the web, desktop or phone app, Free plan included: Add Grill to Claude. Check the name says Grill and the URL says https://grillyour.ai/mcp, then Add and Connect, and sign in. Link didn't open? Customize → Connectors → Add custom connector and paste https://grillyour.ai/mcp. Asking Claude in a chat to connect it does not work.
  • ChatGPT. On chatgpt.com on a computer, then it works in the phone app too: Plugins → + → Add custom MCP server. If that option is missing, turn on Developer mode in Settings → Security and login. Name Grill, server URL https://grillyour.ai/mcp, choose OAuth. Tick I understand and want to continue, then Create, then Install, and sign in.
  • Grok Bot, verified working. Tell the bot Add this MCP server: https://grillyour.ai/mcp, then sign in when asked.
  • Gemini: Should work, testing now. US only, a personal Google account, Keep Activity on. On gemini.google.com: Settings, Connected Apps, Add a custom app, paste https://grillyour.ai/mcp, Next, then sign in.
  • Meta AI (Muse): setup steps are below. Not yet verified end to end, so the site still lists it as coming soon.

Setup for the owner is in docs/HOSTED.md.

Claude Desktop (Mac or Windows): the one-click judge, about 2 minutes with a key, about 10 if you need one

  1. Get a key for Grill's model router: go to openrouter.ai/keys, sign in, click Create key, and add credit.
  2. Install Grill. Download grill.mcpb and double-click it, or drag it onto the Claude window. Paste your key when Claude asks. There's nothing else to install; Claude Desktop runs it.
  3. Try it. In any chat, type: Grill this: we're moving our launch to March. I'm 70% sure it gets us more signups.

The extension reads GRILL_API_KEY. OPENROUTER_API_KEY works when that is unset.

claude.ai on the web or your phone: no key needed

  1. Download grill-skill.zip. In Claude, open Customize → Skills, upload it, and switch it on. To add everything at once instead, go to Customize → Plugins → Add → Add marketplace and enter mtangoz/grill.
  2. Say "grill this: …". Claude writes the judge's prompt. Tap Open in ChatGPT, or paste it into Gemini, then paste the answer back into Claude.

Want it in one step? Use Claude Desktop, above.

ChatGPT, Copilot, DeepSeek, Gemini, Grok or Muse: no install, no key

  1. Copy the Grill prompt into the assistant you think with, and say what you're deciding.
  2. It writes your decision up with you, then gives you a prompt for the judge.
  3. Paste that into an assistant from a different company, then paste the answer back.
You think withJudge with
ChatGPTClaude, Gemini or Grok
CopilotGemini, because Copilot can run OpenAI, Anthropic or xAI models
DeepSeekClaude, ChatGPT, Gemini or Grok
GeminiClaude, ChatGPT or Grok
Grok or GrokBotClaude, ChatGPT or Gemini
MuseClaude, ChatGPT or Gemini

Meta AI (Muse)

  1. In Muse, ask it to add a custom connector named Grill (Meta's help: custom connectors are created in chat; Settings → Connectors is for the built-in list).
  2. URL: https://grillyour.ai/mcp. Transport: Streamable HTTP. Sign-in: OAuth.
  3. If Muse asks for a Client ID, use RNdHSHmWCRJO40vd. Same for everyone. No secret.
  4. Tap Add, then sign in with the same email you signed up with, every time. We email you a code; check spam if it isn't there in a minute.

Cursor, VS Code, Codex, Gemini CLI, or any app that runs MCP servers

bash
npx -y grillyour --set-key      # once: paste your key, Enter, then Ctrl-D

Then add a server named grill that runs npx -y grillyour. The config holds no key. Each app's config, and one-click installs for Cursor and VS Code, are in docs/ANY-APP.md. A Grill Pro key works here too.

Claude Code

/plugin marketplace add mtangoz/grill/plugin install grill@grill

It asks for your model router key. Grill reads GRILL_API_KEY first. OPENROUTER_API_KEY works when that is unset. Or save the key once for every app on this computer with npx -y grillyour --set-key. It goes in ~/.config/grill/key, outside any project's .env.

When Claude Code shows you a plan to approve, Grill adds one line, once a session: want an outside judge on this plan first? Keep planning and say "grill this plan". Claude Code hands the hook the plan, and the hook ignores it: it sends nothing and keeps nothing of the plan. To ask once, it keeps an empty marker file per session in the plugin's data folder, and removes markers older than two days the next time it runs. Turn it off with GRILL_PLAN_OFFER=off.

Claude Code's advisor (/advisor) has a stronger Claude look over the work. That is a second opinion from the same company. Grill's judge is from a different one. Use both.

What it costs

Grill is free. On your own key you pay the router. Copy and paste uses the assistants you already have.

Coming later: an optional Pro plan with saved decision history and look-back reminders. The free tool stays free and account-free. Hear when it's ready.

Privacy

  • Your notes stay in your tools. Only the write-up you approve is sent. The reflection you write stays in your chat or your own notes.
  • The free tool has no account. Grill's makers never see your decisions on that tool. Anonymous usage stats are off unless you turn them on. An in-chat account is open to anyone who signs up; what it stores is in docs/DATA-LEDGER.md.
  • A key, token or password in the write-up stops the run. Email addresses, phone numbers, card and bank numbers, ID numbers, labelled dates of birth, street addresses, postcodes and IP addresses are masked. Names are not. The full list is in docs/PRIVACY.md.
  • The quality check (Jev) is on by default and also sees the masked write-up. Say "skip the check" to skip it for one grill, or switch it off in Grill's settings. The website counts visits without cookies.
  • Grill news line: after a finished grill, at most once a session, and only within 7 days of a release or on the first 3 days of a month, a line with a link to hear when an optional Pro plan is ready. Showing the line sends nothing. Hide it with the "Show Grill news" setting, or GRILL_NEWS=off. It stops on its own.
  • Anonymous usage stats: off by default. One metadata ping to Grill's website after the first successful grill, and never again. Never the decision, the question, the verdict, or your key.

What we will and won't do with your data, and how we build: docs/PRINCIPLES.md.

Turn anonymous usage stats on

Leave them off and the tool does not call Grill. To opt in:

  • Claude Desktop: in the install dialog, switch on "Send anonymous usage stats (optional)".
  • Claude Code: /plugin → Configure options, the same switch.
  • stdio, npx or Docker: set GRILL_USAGE_STATS=on.

Turn them off in the same place, or set GRILL_PING=off. DO_NOT_TRACK=1 also disables the ping. The paste route does not send it.

Details: docs/PRIVACY.md.

Weekly review

Optional, in Claude, when you keep your own decision log: docs/WEEKLY-REVIEW.md.

License

MIT. See LICENSE. The code that handles your decision stays MIT for good (docs/PRINCIPLES.md). The name and logo are not part of the license: TRADEMARKS.md.

来源:README.md,提交 ede08ab

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.1.0最新Sep 27, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分