Sahawatthanakit (1988) Co., Ltd.
SAHAWATTHANAKIT(1988) · Make It Smart
Free · Chrome extension + Claude Desktop · ChatGPT · Claude · Gemini

AI Guard — make the AI finish the job before you send it

One button. Every number, name, date, law and reference in the answer goes back to the AI, which must quote your source or mark it unverified — and then rewrite the final version.

The problem it removes

AI answers look finished. The date on the quotation, the gross-profit figure, the section of law, the standard number — any of them can be invented, and nothing on the page tells you which. Checking by hand takes longer than writing it yourself, so most people send it as is.

AI Guard does not judge right or wrong — it cannot. It finds the kind of fact AI gets wrong most often, sends the list back, and makes the AI do the checking against the material you actually gave it. What you get back is the same document with the guesses marked or removed.

  1. 1

    Ask as usual

    Chat with ChatGPT, Claude or Gemini the way you always do. The extension adds one small bar at the bottom of the page.

  2. 2

    Press “Finish → ready to send”

    The scanner (plain rules, no AI) lists every money figure, quantity, date, person, company, standard, law and reference in the last answer, and pastes them back into the chat.

  3. 3

    The AI closes each item

    It must quote the exact source you gave, redo the arithmetic, or mark the claim [verify] — then it rewrites the final version with nothing invented left inside.

Measured, not promised

94

real chats it was run on (5 Sep 2026: Claude 30 · ChatGPT 28 · Gemini 36)

41

claims the AI admitted it had guessed — a wrong quote date, a wrong gross-profit figure, a made-up price validity period

0

times the AI changed a correct figure into a wrong one

−24%

false alarms removed in v0.3.2 after a second pass over 92 of those chats (605 → 460 flagged items)

Both runs were done on the owner’s own chats, by the same team that built the tool. Treat them as a first measurement, not an independent benchmark. The scanner is plain pattern rules: it will still flag some things that are fine (a product code that looks like a unit, a common word after “Company”), and it cannot see a guess that carries no number, name or reference.

Install

Chrome extension

Download v0.4.3 (zip, free)

Chrome Web Store listing is in preparation. Until it is live, install from the zip in four steps — no account, no sign-in.

  1. Download the zip (v0.4.3) and unzip it.
  2. Open chrome://extensions, turn on “Developer mode” (top right).
  3. Click “Load unpacked” and choose the unzipped folder (ai-guard-extension-0.4.3 — the folder that contains manifest.json).
  4. Open chatgpt.com, claude.ai or gemini.google.com — the bar appears at the bottom. Switch TH/EN with the button on the bar.

Claude Desktop extension

Download v0.4.7 (.mcpb, free)
  1. Download ai-guard-0.4.7.mcpb.
  2. Double-click the .mcpb file — Claude Desktop shows an install dialog.
  3. Click “Install”.
  4. Open Settings → Extensions in Claude Desktop — AI Guard is listed there.

Privacy policy — AI Guard

Effective date: 14 September 2026

This policy covers the AI Guard Chrome extension (v0.4.3) and the AI Guard extension for Claude Desktop (v0.4.7), published by Sahawatthanakit (1988) Co., Ltd. — we run no server for AI Guard: neither product sends anything to us, and neither has an account, analytics or telemetry.

Chrome extension

  • Runs only on chatgpt.com, chat.openai.com, claude.ai and gemini.google.com (the content-script match list in the manifest).
  • Permissions: “storage” and “clipboardWrite” only. It does not request “tabs”, “history” or any host permission, and it makes no network request.
  • It reads the AI answer shown on the page and, when you press a button, writes text into the chat box or copies it to your clipboard. Text in the chat box goes to ChatGPT, Claude or Gemini only when you send it yourself, under that service’s own privacy policy.
  • Auto-check (on by default, “Auto: on/off” button on the bar): when an AI answer has stopped changing for a few seconds, the extension scans that answer once, on the page, with the same local scanner — then shows a label under the answer (“AI Guard checked · N to verify”) and a number on its toolbar icon. It never edits the answer and sends nothing anywhere.
  • A background script only animates the extension’s toolbar icon and sets that number on it; it adds no permission and makes no network request.

Claude Desktop extension

  • A local program that Claude Desktop starts on your computer. Claude passes text to its tools (task_pack, recompute, scan_claims, finish_answer, check_quotes, guard_status) and gets text back. Your conversation with Claude itself is handled by Anthropic under Anthropic’s own policy.
  • task_pack, recompute, scan_claims, finish_answer and guard_status run only on your computer and make no network request.
  • Notification when AI Guard runs (on by default; turn it off in Claude Desktop → Settings → Extensions → AI Guard): on macOS, a system notification appears when AI Guard really starts a task and when it checks the answer before sending, showing the task and counts, e.g. “Finish check · 11 claim(s) · 5 calculation(s) · id BEBBEA”. It never contains your text, numbers or links. macOS keeps delivered notifications in Notification Center under its own settings. The extension may also send progress messages to Claude Desktop while a tool runs.
  • check_quotes is the only tool that goes online. For each link the answer cites as [SOURCE: link | “quote”] (at most 20 different links per call) it sends one plain HTTP GET request for that page, directly from your computer, then checks whether the quoted sentence appears in the page text. The request contains the page address and standard request headers (user-agent “Mozilla/5.0 (compatible; AI-Guard/<version>; quote check)”) — it does not send the answer, the quote or any personal data. As with any web request, the site receiving it can see your IP address.
  • check_quotes opens only http/https links. It refuses local and private-network addresses (localhost, .local, .internal, 127.x, 10.x, 172.16–31.x, 192.168.x, 169.254.x, 100.64–127.x, 0.0.0.0, IPv6 loopback, private and link-local) and checks every redirect again (at most 5). Each page has a 15-second timeout and an 8 MB limit, and only text / HTML / XML pages are read.

What is collected

  • By us: nothing. No personal data, chat content or usage statistics reach us from either product.
  • Stored in your browser by the Chrome extension (chrome.storage.local): your language choice (TH/EN); whether auto-check is on or off; the last task pack you picked (its name only, not your source); and your test log — only the items you mark yourself as hit / missed / noise, each with up to 300 characters of the marked text, the scanner category, the site name and the time.
  • Stored on your computer by the Claude Desktop extension, in the file ~/.ai-guard/usage.json (or the folder set in AI_GUARD_HOME): how many times each tool ran on each day; the name and time of the last run; and, for finish_answer and check_quotes, the 6-character stamp id, tool name, time and number of items checked. It does not store answers, your source text, quotes, links or page content.

How it is used and stored

  • Chrome: the language and task choices only restore your settings. The test log exists so you can copy it (“Copy test results”) and send it to us if you choose — nothing is sent automatically.
  • Claude Desktop: the usage file lets guard_status tell you whether AI Guard ran today and whether a stamp shown at the end of an answer was really issued on this computer.
  • Pages opened by check_quotes, and text passed to any tool, are processed in memory only for that tool call and are not written to disk.

Third-party sharing

  • We do not sell, rent or share any data — we receive none to share.
  • The only traffic that leaves your device because of AI Guard is the check_quotes page requests described above, which go to the sites linked in the answer. Text you choose to send in ChatGPT, Claude, Gemini or Claude Desktop goes to that AI provider, as it would without AI Guard.

Retention

  • Chrome: stored settings and the test log stay in your browser until you press “Clear log” in the test-results panel or remove the extension.
  • Claude Desktop: the usage file stays on your computer. Day counts older than the latest 90 days and stamps beyond the latest 500 are removed automatically. The extension never deletes the file itself — delete the ~/.ai-guard folder to remove it at any time.
  • check_quotes page content is discarded when the tool call ends. We retain nothing, because we receive nothing. If you email us your test log, that email is handled under our company privacy policy (PDPA) linked below.

Contact

Questions or requests about this policy or either extension:

Sahawatthanakit (1988) Co., Ltd. · sales@sahawatthanakit1988.com

Company privacy policy (PDPA)

Tell us what it missed

The bar has a “Test results” button that copies your hit / missed / noise log as text. Send it to us and the next version is built from it — that is exactly how v0.3.2 came out of v0.3.1.

sales@sahawatthanakit1988.com