Check an AI answer, claim by claim.

Paste a passage from ChatGPT, Claude, or Gemini. Each factual claim is pulled out, checked against established sources, and underlined inline.

Checks: 0
Claims: 0
Verified: 0%

Text to check 0 words

Examples:

0 / 20,000 Batch mode

1

Extract atomic claims

Splits your passage into single factual statements, skipping opinions and filler while keeping numbers and dates intact.

2

Verify without guessing

Checks each statement against established records. Post-cutoff or uncertain claims are marked unverifiable rather than guessed.

3

Inline proof & links

Underlines the original passage in place and attaches a correction and primary link to every claim.

No checks yet

Run a check and it will show up here.

Most cited

Nothing cited yet

Sources appear here automatically once checks are run.

Reference directory

Public records, journals, and encyclopedias prioritized when checking claims.

Public & International

Medicine & Health

Science & Computing

Encyclopedias & Wire Services

Endpoints & Package

Agent Skill CLI npx skills add Sunil56224972/hallucination-Hunter

Remote MCP (HTTP) https://hallucination-hunter-five.vercel.app/api/mcp

Verify & DB API https://hallucination-hunter-five.vercel.app/api/verify

ChatGPT Schema https://hallucination-hunter-five.vercel.app/openapi.json

Setup by Platform

1. Install the Agent Skill

Adds hallucination-hunter to Antigravity so the agent verifies factual claims in its response, corrects false statements, and logs checks to Supabase.

# Install via Agent Skills CLI
npx skills add Sunil56224972/hallucination-Hunter

# Or clone directly into Antigravity skills directory
git clone https://github.com/Sunil56224972/hallucination-Hunter.git
cp -r hallucination-Hunter/skills/hallucination-hunter ~/.gemini/config/skills/

2. Optional: Add the MCP Server

Exposes native verify_llm_response and search_factcheck_database tools inside Antigravity’s MCP config.

{
  "mcpServers": {
    "hallucination-hunter": {
      "url": "https://hallucination-hunter-five.vercel.app/api/mcp"
    }
  }
}

1. Claude.ai (Web & Desktop Skills)

Download hallucination-hunter-skill.zip, open Claude Settings → Capabilities → Skills, and upload the zip. For Claude Code CLI, run:

npx skills add Sunil56224972/hallucination-Hunter -a claude-code

2. Claude Desktop MCP Config

Add to claude_desktop_config.json (remote HTTP or local node mcp/server.mjs) so Claude checks drafts and queries the 900+ claim database.

{
  "mcpServers": {
    "hallucination-hunter": {
      "command": "node",
      "args": ["/path/to/hallucination-Hunter/mcp/server.mjs"],
      "env": { "HH_CLIENT": "claude" }
    }
  }
}

1. Import Action Schema URL

In ChatGPT, open Explore GPTs → Create → Configure → Actions → Import from URL and paste this OpenAPI 3.1 endpoint (Authentication: None):

https://hallucination-hunter-five.vercel.app/openapi.json

2. Paste Custom GPT Instructions

Instructs ChatGPT to verify its factual answers against Hallucination Hunter’s engine + database and display the audit block beneath its reply.

When answering any prompt with objective factual claims (dates, numbers, versions, names, scientific or historical facts):
1. Draft a clear, direct answer first.
2. Call verifyLlmResponse with your drafted text and client="chatgpt".
3. If any claim returns status="false", fix that statement in your answer using the returned correction.
4. Present your final answer first, then append the returned markdownReport block underneath so the user gets both your answer and the database-synced Hallucination Hunter verification.

1. Verify Any LLM Output & Sync to Database

Works with Gemini, Cursor, Windsurf, Codex CLI, LangChain, LlamaIndex, Dify, OpenWebUI, or n8n HTTP nodes. Pass optional Authorization: Bearer gsk_... for dedicated throughput.

curl -sS -X POST https://hallucination-hunter-five.vercel.app/api/verify \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Python was created by Guido van Rossum in 1991. Python 3.0 shipped in 2005.",
    "client": "gemini-agent",
    "saveToDatabase": true
  }'

2. Query the Shared Supabase Fact-Check Database

Search previously verified or contradicted claims across all connected assistants directly via GET /api/verify.

# Search stored claims by keyword & status
curl -sS "https://hallucination-hunter-five.vercel.app/api/verify?search=Python&limit=5"

# Retrieve live database totals and recent cross-LLM checks
curl -sS "https://hallucination-hunter-five.vercel.app/api/verify?limit=5"

Live Co-Pilot Test

Runs POST /api/verify and syncs to Supabase

Simulated client

Returned markdownReport (appended beneath the LLM’s answer)

Preferences

ModelUsed to extract and verify factual claims.

Confidence thresholdClaims below this confidence are dimmed.

75%

ThemeStored in this browser.

Keyboard shortcutsPress ? anywhere outside a text field.

Database

Supabase PostgresStores past checks, extracted claims, and settings.

Checking…

Saved checksTotal analyses stored in Postgres.

–

Saved claimsIndividual statements checked across all runs.

–

RegionDatabase host location.

ap-south-1 (Mumbai)

Language models write in the same confident tone whether they are stating a well-known fact or inventing a number. Hallucination Hunter breaks an answer into single claims so you can see which lines hold up and which need a second look.

How a check runs

Split into claims. Your text is broken into standalone factual statements, keeping exact numbers, dates, versions, and names intact.
Cross-check & verify. Each statement is checked against established knowledge, live DuckDuckGo source lookup, and prior records in the Supabase claims ledger. Post-cutoff or unconfirmed claims are marked unverifiable rather than guessed.
Attach sources & corrections. Contradicted claims include a one-sentence correction and a source URL you can open yourself.
Underline inline. Claims are marked directly inside your original passage so errors stand out immediately.

Works inside your LLM

Install the bundled Agent Skill (npx skills add Sunil56224972/hallucination-Hunter), attach the remote MCP server (/api/mcp), or import the OpenAPI 3.1 schema (/openapi.json) into ChatGPT Actions. Connected assistants return their normal answer alongside a Hallucination Hunter verification block and sync every check back to the shared Supabase Postgres ledger.

Privacy

Checked text runs through serverless Vercel endpoints to Groq and is stored in Supabase Postgres under a random browser ID or assistant client tag (pass saveToDatabase: false for ephemeral checks). There is no account wall, tracking cookie, or analytics script.