Six AI detectors.
One POST. No API key.

Send any text, get back AI-generation scores from ZeroGPT, Grammarly, Copyleaks, Originality.ai, Pangram and QuillBot — each provider independently, plus a consensus average. Free, unmetered, no API key.

Copy a snippet

// Quickstart

Everything hangs off a single endpoint. Set $BASE to this server’s origin: this page’s URL.

curl -s "$BASE/v1/detect" \
  -H "Content-Type: application/json" \
  -d '{"text": "Paste the text you want checked here. Detectors generally need a paragraph or two; Originality.ai requires at least 100 words."}' | jq
const res = await fetch(`${"$BASE"}/v1/detect`, {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ text: "Your text here…" }),
});
const data = await res.json();
console.log(data.aggregate.average_ai_score_percent);
// -> 87.4  (consensus AI score across providers)
import requests

r = requests.post("$BASE/v1/detect",
                  json={"text": "Your text here…"})
scores = r.json()["scores"]
for provider, s in scores.items():
    print(provider, s.get("ai_score_percent"), s.get("verdict"))

// Endpoints

POST/v1/detect

Run the gauntlet. Body: {"text": "...", "providers": ["optional","subset"]}. Returns every provider’s score plus the consensus average. Timeout-tolerant: a slow provider never blocks the others.

GET/v1/detect?text=…

Same result in one URL — handy from a browser. Optional &providers=zerogpt,grammarly.

GET/v1/providers

Machine-readable list of detectors with their limits and quirks.

GET/health

Liveness + pool depth + provider roster.

GET/openapi-ui · /openapi.json

Interactive Swagger UI and the raw OpenAPI 3 schema.

// Providers

idsourcerangenotes
zerogptzerogpt.comany lengthreports AI % of words + detected language
grammarlygrammarly.com1 – 10k wordscharacter ranges flagged as AI included
copyleakscopyleaks.com1 – ~5k wordsscans in ~100-word chunks, word-weighted average; short texts need ≥150 characters
originalityoriginality.ai≥100 – 50k wordstexts under 100 words come back skipped, not failed
pangrampangram.comany lengthchunked at 100 words, per-chunk predictions reported
quillbotquillbot.com1 – 30k charssentence-level chunk breakdown included
Skipped ≠ failed. A provider that can’t legally take your text (too few words, etc.) returns "status": "skipped" with an explanation in "note". The consensus average uses only "ok" providers.

// Response shape

{
  "text_characters": 1210,
  "text_words": 214,
  "duration_ms": 24311,
  "aggregate": {
    "average_ai_score_percent": 63.17,
    "verdict": "ai",
    "providers_scored": 6,
    "providers_skipped": 0,
    "providers_failed": 0
  },
  "scores": {
    "zerogpt": {
      "status": "ok",
      "ai_score_percent": 74.5,
      "human_score_percent": 25.5,
      "verdict": "ai",
      "latency_ms": 4120,
      "detail": { "ai_words": 118, "total_words": 159, "detected_language": "English" }
    },
    "grammarly":  { "status": "ok", "ai_score_percent": 62.0, "...": "..." },
    "copyleaks":  { "status": "ok", "ai_score_percent": 51.3, "...": "..." },
    "originality":{ "status": "ok", "ai_score_percent": 71.2, "...": "..." },
    "pangram":    { "status": "ok", "ai_score_percent": 55.8, "...": "..." },
    "quillbot":   { "status": "ok", "ai_score_percent": 64.2, "...": "..." }
  }
}
aggregate.average_ai_score_percent
unweighted mean of every ok provider’s AI score. null if nothing scored.
verdict
human ≤25% · mixed 25–50% · ai >50%
scores.*.status
ok · skipped (input didn’t meet that provider’s rules) · error (upstream hiccup — retry freely)
scores.*.detail
whatever extra intel that provider returned: flagged char ranges, per-chunk scores, models…
Reading the number: every ai_score_percent is “probability this text is machine-generated” from that vendor. Different vendors genuinely disagree — that disagreement is the signal. 3+ agreeing is strong; outliers are usually the noise.

// Limits & fair use

  • No key, no signup, no rate limit. Scan as much as you want — usage is unmetered.
  • 60,000 characters per request max.
  • Heavy bursts just queue momentarily (concurrency is pooled) — you effectively always get served.
  • Scores are unofficial queries of each vendor’s public web tools — treat them as signals, not verdicts.
  • Don’t build an academic-punishment case on one vendor alone; triangulate.
  • Cache results for identical text when you can — it keeps the shared infrastructure snappy for everyone, including you.

// Self-hosting

The whole thing is one small Docker image. The rotating residential proxy pool ships inside the repo (data/proxies.txt) — nothing to bring:

docker run -d -p 8000:8000 \
  -v $(pwd)/data:/app/data \
  polygraph-api

# data/proxies.txt            bundled rotating-residential pool (49k+ identities)
# data/copyleaks_tokens.json  token cache (auto-created, kept alive by the server)
# optionally: ANYSOLVER_API_KEY / GRAMMARLY_GRAUTH / … env overrides

Source layout: aggregator_app/ (FastAPI + adapters), data/ (runtime state). Deployable to Coolify/Dokploy/K8s as a plain Dockerfile service exposing port 8000.