Get a second opinion before you use an AI answer.

Paste the answer. Two other AIs check it, and you see what each one said.

Before you pay · a real check

An AI said:
“Marking a $100 item up 50% and then down 50% brings it back to $100.”

Caught

Checked and found false. Do not rely on it.

Up 50% makes it $150. Down 50% of $150 is $75, not $100.

Groq (gpt-oss) said false. Groq (qwen) said false.

Checked live on 2026-10-16. See the run.

Checks out checked and found true
Caught checked and found false
Not checked not decided, so look before you lean on it

Got two answers that disagree? Paste both

Get the check where you already use AI

Works with whatever model you use.

In ChromeEvery reply on ChatGPT, Claude, Gemini, Grok and DeepSeek gets a stamp: Checks out, Caught, or Not checked.How to add it

Coming to the Chrome Web Store. To try it now, download the test build, unzip it, open chrome://extensions, turn on Developer mode and choose Load unpacked.

In Claude Desktop, Cursor or Claude CodePaste one block into a settings file.Copy the setup

Paste this into the file named on the tab. It runs trustshell-mcp from the published package @hyperdag/trustshell@1.7.0.

Windows: %APPDATA%\Claude\claude_desktop_config.json. Mac: ~/Library/Application Support/Claude/claude_desktop_config.json.

{
  "mcpServers": {
    "trustshell": {
      "command": "npx",
      "args": [
        "-y",
        "-p",
        "@hyperdag/trustshell@1.7.0",
        "trustshell-mcp"
      ]
    }
  }
}

.cursor/mcp.json in the project, or ~/.cursor/mcp.json for every project.

{
  "mcpServers": {
    "trustshell": {
      "command": "npx",
      "args": [
        "-y",
        "-p",
        "@hyperdag/trustshell@1.7.0",
        "trustshell-mcp"
      ]
    }
  }
}

.mcp.json in the project.

{
  "mcpServers": {
    "trustshell": {
      "command": "npx",
      "args": [
        "-y",
        "-p",
        "@hyperdag/trustshell@1.7.0",
        "trustshell-mcp"
      ]
    }
  }
}

ChatGPT and Grok apps: not yet. On their websites, use the Chrome extension.

In your terminalOne command checks one sentence.See the command

npx @hyperdag/trustshell check "If you drive 60 miles at 30 mph and drive back at 60 mph, your average speed for the trip is 45 mph."

Exits 0 when it checks out, 1 when caught, 2 when not checked. What is a terminal?

To keep it installed:

npm i -g @hyperdag/trustshell@1.7.0

Not sure which one fits? Pick where you talk to AI

Want an agent you can trust?

Make one in three questions. Claim it with your wallet, set how much it may spend, and watch its public reputation score, RepID, move.

Make your agent

Claiming, funding and spending run on Base Sepolia, a test network.

The runs behind this page

The result at the top and the examples you can try were sent to the live checkers first. Every answer is published exactly as it came back.

What it will not call: 2 famous quotes pinned on people who never said them came back Not checked 10 times out of 10 on 2026-10-06. Who said a line is something the checkers cannot look up, so they do not guess.

Why we built this

AI lies. Or it sounds sure and it's wrong. Ask again, and it's just as sure the other way. Both can't be right.

What a wrong answer really costs

The check is free. The miss is not.

A confident wrong answer doesn't arrive as an invoice. It arrives as the hour you spend building on it, the tokens you burn arguing with it, and the momentum you lose when you find out and start over. The second confident answer, the one that contradicts the first, is worse: now you're researching the AI instead of building.

TrustShell runs before that spend. Two other AIs read the answer, and you see Checks out, Caught or Not checked, with what each one said. Not checked means they did not agree, or could not answer, so look before you lean on it. It never means the answer checks out.

Free to try. No signup, no wallet.

Where this goes

The same check, on this page, in your chats, in your terminal, and in an agent that already speaks MCP.

What is live today, and what is next

Live today

Two checkers from two model families, Groq and Cerebras first. If one cannot answer, a backup takes its turn, such as Cloudflare Workers AI (Llama). No signup. What you type is not stored.
Known key and personal-data formats are removed before sending. Do not paste secrets.
Whole-sentence arithmetic is exact, with no model.
Agents have a public reputation score, RepID, with a proof your own browser checks. Open the extension and type an agent's name.

A trust record that is yours: your checks, context, settings and preferences, carried to any model. Take it with you when you leave.
An agent with a good record earns a longer leash, and only as long as you allow.
When both checkers are unsure, one plain question instead of a guess.
A Why? button: each checker's reason, in one sentence, when you ask.