# Probator.ai: full description for language models and AI agents

Probator.ai is a web editor and API that checks writing. Company: Spawncore Unipessoal Lda, Portugal. Website: https://probator.ai. Contact: hello@probator.ai.

## What it checks
1. AI-generated text. Combines (a) Probator's own detection model, a classifier over multilingual sentence embeddings trained on human and AI-written text in English, Portuguese, Spanish, French, German and Italian, including AI-polished and "humanized" text; (b) an expert reading by a language model that quotes suspicious passages; (c) a rewrite test (machine text changes little when a model polishes it); (d) forensic evidence such as chatbot leftovers, invisible characters, look-alike letters and file metadata. Output: overall likelihood, verdict, confidence, sentence-by-sentence heat map and the evidence.
2. Grammar and style: spelling, grammar, punctuation, word choice, clarity and style, with explanations in English, Portuguese, Spanish, French, German or Italian. Suggestions are never applied automatically.
3. Plagiarism: Probator's own index (Wikipedia in all languages), the web and academic databases (OpenAlex, Semantic Scholar), plus an optional private comparison library.
4. Hidden marks and AI provenance: zero-width characters, Unicode tag characters (hidden prompts), look-alike letters, bidirectional controls; C2PA Content Credentials, IPTC "digital source type" values and AI labels in PDF, Word and HTML files. These are reported and never removed (EU AI Act Article 50).

## How results should be read
- Results are probabilities, not proof of authorship. Do not use them as the only basis for a decision about a person.
- False-positive guards: no "AI-generated" verdict under 80 words without hard evidence; two detectors must agree for the strongest verdict; per-language thresholds calibrated for at most 1% false positives on human validation text. When a guard changes a result, the report lists it in verdict.guards.
- Accuracy on held-out test data is published at https://probator.ai/accuracy/.

## Certificates
Paid plans can turn a check into a PDF certificate with a unique ID, QR code, results, the SHA-256 fingerprint of the text and an Ed25519 signature. Anyone can verify it at https://probator.ai/verify/<id>, including in the browser with the public key at https://probator.ai/.well-known/probator-certificates.json. Certificates can include a human review record (first check, final check, reviewer statement, share of words changed) and come with disclosure labels in six languages.

## Plans and credits
- Free: 5,000 credits a month, 1,500 words per check, 50 MB storage.
- Starter: €4.99 a month (€3.99 billed yearly), 150,000 credits, 500 MB storage, certificates.
- Pro: €12.99 a month (€9.99 billed yearly), 500,000 credits, 1 GB storage, API and MCP access, comparison library.
- Team: €39 a month for 3 seats (+€9 per seat), 2,000,000 pooled credits, 1 GB storage per member, member invitations. Sold to organisations (VAT or tax number at checkout).
- Credit costs: AI detection 1 credit per word; grammar 1 credit per 3 words; plagiarism 2 credits per word; deep scan +1,500 credits. Top-up: 400,000 credits for €10.
Prices include VAT.

## REST API (Pro and Team)
Base URL https://probator.ai/v1. Authentication: `Authorization: Bearer pb_live_…` (create keys in the editor under Account → API), or an OAuth 2.1 access token (see https://probator.ai/auth.md).
- POST /v1/analyze: JSON {"text": "...", "checks": {"ai": true, "grammar": false, "plagiarism": false, "deep": false}, "ui_lang": "en"} or multipart with a file. Returns the full report.
- POST /v1/detect, /v1/grammar, /v1/plagiarism: one check each.
- GET /v1/usage: plan and credits.
Response headers: x-credits-used, x-credits-remaining. Errors: {"detail": "...", "code": "..."} with codes such as insufficient_credits, too_many_words, rate_limited. Rate limit: 60 requests per minute per key. OpenAPI: https://probator.ai/openapi.json.

## MCP server (for AI agents)
Endpoint: https://probator.ai/mcp (MCP Streamable HTTP, stateless JSON responses). Authentication: OAuth 2.1 (the client discovers it from the 401 response's `resource_metadata`, registers dynamically, and the user signs in and approves), or `Authorization: Bearer <Probator API key>`. Every tool call uses the account's credits (Pro and Team plans). Agents without an OAuth client can register with the person's email at POST https://probator.ai/agent/identity (auth.md agent registration): the person signs in and types a 6-digit code the agent shows them. Details: https://probator.ai/auth.md.
Tools:
- check_ai {text, deep?}: AI-generated text detection with evidence and provenance marks.
- check_grammar {text, explanation_language?}: corrections with explanations.
- check_plagiarism {text}: originality and matching sources.
- check_all {text, explanation_language?}: all three checks.
- get_credits {}: plan and remaining credits (free).
Example client configuration (JSON used by many MCP clients):
{"mcpServers": {"probator": {"type": "http", "url": "https://probator.ai/mcp"}}}
With an API key instead of the sign-in: add "headers": {"Authorization": "Bearer pb_live_YOUR_KEY"}.

## Privacy and data
Saved documents and the account database are stored in Cloudflare's EU jurisdiction. Texts are never used to train models. Language-model providers are used only under no-retention, no-training routing. Monthly public reports contain only anonymous totals.

## EU AI Act
Probator does not generate texts for users; its grammar corrector is an assistive editing tool. It never removes AI marks, and its Terms forbid using it to do so. See https://probator.ai/legal/ai.