See through synthetic text.
A desktop app in development for local text detection and rewriting. Compose and optional cloud humanization send the submitted text to Claude; detection can work offline. MCP workflows follow the same operation-specific boundaries.
AI text is everywhere
ChatGPT, Claude, Gemini. Everyone uses them now. But AI-generated text has a signature: uniform sentence lengths, predictable transitions, hedging language, corporate jargon. No personality, no voice, no humanity.
Writers need to know if their drafts sound robotic. Editors need to spot AI-assisted submissions. Professionals need to polish AI-drafted content without sounding like a machine.
Detection that stays on your machine
PRISM analyzes text locally using heuristics and machine learning, then shows scores and sentence-level signals. These are fallible indicators, not proof of authorship or a basis for consequential decisions on their own.
Humanization offers local ONNX rewriting when the required models are installed, or Claude API rewriting in the configured cloud mode. Cloud humanization and Compose transmit the submitted text to Anthropic and require an API key.
System architecture
PRISM is built with Tauri 2.0 — a lightweight Rust-based desktop framework. The React frontend handles the UI and state management through Zustand, while all analysis runs natively in Rust.
Text analysis combines ONNX-based machine learning inference with heuristic metrics: perplexity, burstiness, entropy, and pattern recognition. The Claude API provides humanization and compose capabilities when configured.
API keys use the OS keychain. Detection, local rewriting, and exports can stay on the device with the required models installed. Compose and cloud humanization call the Claude API, including when invoked through MCP; those operations send text outside the device.
Two modes
Detect
Analyze and humanize
- 1Paste text
- 2View AI probability score and metrics
- 3See heatmap highlighting AI-like sentences
- 4Click humanize for 3 style variations
- 5Export results as JSON or PDF
Compose
Draft and polish through Claude
- 1Write rough text in any language
- 2Select source and target language
- 3Get 3 polished versions: Natural, Creative, Professional
- 4Copy the one you like
- 5Repeat and refine
In practice
Detect mode with heatmap analysis, and Compose mode for polishing drafts.
Detect mode — AI probability score with sentence heatmap
Detect mode — detailed metrics and pattern analysis
Compose mode — three polished variations from rough text
Compose mode — multi-language support with style presets
What PRISM includes
Detection, humanization, composition, and export in a single desktop application.
Local ML detection (ONNX)
Trained classifier runs locally via ONNX Runtime. No internet required for detection.
Heuristic analysis
Perplexity, burstiness, and entropy metrics reveal AI writing patterns.
Pattern recognition
Detects filler words, hedging language, corporate jargon, and uniform sentence structures.
Sentence-level heatmap
Color-coded visualization shows which sentences triggered detection and why.
Three humanization styles
Natural, Creative, and Professional. Each rewrites AI text with a distinct voice.
Experimental local rewriting
Local ONNX rewriting with the required models installed. Quality varies and needs human review; it is not equivalent to the Claude cloud mode.
Compose mode
Send a draft to the Claude API with a source and target language to receive style variations. Requires an API key and transmits the text to the provider.
Multi-language support
Detection and humanization in English, Italian, Spanish, and German.
PDF and JSON export
Export full analysis reports as styled PDFs or structured JSON data.
Secure API key storage
Anthropic API key stored in your OS keychain. Never in config files, never logged.
What stays local, and what connects
Current implementation boundaries, not a comparison of competitors, prices, or detection accuracy.
| PRISM | |
|---|---|
| Runs locally | Desktop app and detection engine |
| Works offline | Detection and local rewriting with models installed |
| Source license and access | MIT in source; public repository access unavailable |
| Humanization | Local ONNX or optional Claude cloud mode |
| Compose mode | Claude API; text leaves the device |
| Sentence heatmap | Available in the development build |
| Export (PDF/JSON) | Local PDF and JSON exports |
| Price | No verified commercial plan; provider usage may cost extra |
| Data privacy | Local for detection; Claude processes cloud requests |
Stack
Built with Tauri 2.0 and Rust for native performance and small binary size.
- Desktop
- Tauri 2.0
- Frontend
- React 19 + TypeScript + Tailwind
- Analysis engine
- Rust (heuristics + ONNX ML)
- AI integration
- Claude API or local ONNX
- Export
- genpdf (PDF) + serde_json (JSON)
- Languages
- EN, IT, ES, DE
Access and setup
The previous source link is not publicly accessible. PRISM remains in development; contact us about access and setup. This page does not promise an available download or a supported release.
The name
PRISM — PRISM — Pattern Recognition for Identifying Synthetic Messages.
Like a prism splits light into its components, PRISM breaks down text to reveal its true nature. Not anti-AI. Pro authenticity.
Roadmap, not an available service plan
Managed REST APIs, webhooks, custom model training, and support packages are proposals to validate. They are not currently offered here as included product capabilities or purchasable plans.
Cloud operations require your own provider credentials; provider terms and usage charges are separate.
License and source access
The checked source carries the MIT license, but the previous repository link does not provide anonymous public access. Contact us for current access conditions; no source download is offered on this page.
Discuss PRISM.
Ask about the development build, source access, and which operations fit your privacy requirements.
Ask about access