Your stories, your choice

The Three Tiers of Privacy

We believe you should decide where and how your data is processed. Processing preferences control new protected-content actions; account, existing campaigns, collaboration, and content previously saved to the server remain server-backed.

Cloud-First

Processing happens on our servers and configured API providers. This offers the fullest feature set; content you save is stored in your account.

Hybrid

Choose local audio or OCR while allowing selected text features to use server processing. Check each action's disclosure before sending content.

Device Private

Local recap, search, opener, catch-up, and transcription paths do not fall back to cloud. New protected uploads, sync, cloud OCR, Deep Recall, and server enrichment are unavailable while active.

What the AI actually does

Parcane is designed to organize material you provide, not to author your world or decide what happens next. Generated summaries and extracted details can still be mistaken; review them against your source notes before treating them as canon.

The creative work — the world, the characters, the moments that matter — that's all you and your table. We just make sure you don't forget any of it.

Why we're transparent about this

We know AI is a loaded topic in the TTRPG community, and we think that skepticism is healthy. The best tabletop moments come from human creativity, improvisation, and connection — no language model can replace that, and we don't think it should try.

We use AI the same way you might use a spell-checker or a search engine — as invisible infrastructure that saves you time on the boring parts so you can focus on the parts that matter. You'd never call a spell-checker a "co-author," and we don't think organizing your session notes makes us one either.

The technical details

For those who want specifics, here's exactly what happens to your data:

• Text recaps & Narration — You choose the engine. Cloud: Notes are sent to the configured language-model provider (currently Google Gemini) on a paid API tier whose terms prohibit using your content to train models. Local: Notes are processed in your browser using WebLLM or wllama. In Device Private, protected-content save and enrichment actions are blocked rather than uploaded.
• Audio transcription (cloud) — Recordings are sent through our server to the configured provider, currently Groq or OpenAI depending on deployment settings. Temporary processing files are deleted after the request completes.
• Audio transcription (on-device) — When you choose "On This Device," your recording is transcribed entirely in your browser using a local speech model (Whisper via Transformers.js). The audio never leaves your device — no server, no API, no upload. The model weights are stored in your browser's Cache API for future use.
• Local AI Narration — If you select a local engine like WebLLM or wllama, inference runs in your browser's WebGPU or WebAssembly environment. Device Private disables its model recommendation and result telemetry transfers; model assets may still need to be downloaded when you prepare a device.
• Session Search & Opener — When you search or ask for a "Previously On," the chosen engine analyzes available recap context. Device Private forbids cloud model fallback, although previously server-stored campaign context may still be downloaded for local processing.
• Story threads & Catch-up — Catch-up generation follows the recap engine and is locally enforced in Device Private. Server-only enrichment features are disabled in that mode.
• Wiki editing (text & voice) — When you type or speak a correction like "rename the blacksmith to Alvin," the current feature sends the instruction for server-side interpretation and mutation. It is unavailable in Device Private until a local path is implemented.
• Voice routing — The Omni-Mic uses a lightweight language model to classify your spoken input (wiki edit? search query? quick note?) and route it to the correct feature. It categorizes — it doesn't execute anything on its own.
• Handwritten note OCR — Photographed pages are read using optical character recognition. Cloud: multi-model consensus (GPT-4o, Gemini Flash) for highest accuracy. Local: Tesseract.js runs in your browser. OCR output is a draft to review; Device Private disables premium cloud OCR.
• Ambient recording — Long-form background recording for in-person sessions. Audio chunks are stashed in your browser's IndexedDB for crash recovery. Transcription follows your processing choice. Chunks remain locally for recovery until successful cleanup or discard; cloud transcription also uses temporary server processing files.
• Account data — Stored in our database on a dedicated server. Encrypted at rest. Backed up daily. You can request deletion at any time. Device Private blocks new protected-content saves and sync; it does not delete account data or content already stored on our servers.
• How providers handle your content — We use paid API tiers whose terms contractually prohibit training AI models on your inputs or outputs (Google Gemini, OpenAI, Groq). Providers may retain content briefly (typically up to 30 days) solely to run the service and detect abuse, after which it is deleted. To be clear about the distinction: "never used for training" is not the same as "never touched" — cloud content is processed, and briefly held, to deliver the feature you asked for. If you want zero cloud transmission, use the on-device and Device Private options above.

What you can always do

• Delete your data — Remove individual recaps, entire campaigns, or request deletion of your account data.
• Export your data — Download individual campaigns as Markdown or your entire account history as a single JSON bundle. Your data belongs to you.
• Use it without AI — Parcane accepts plain text notes. You can paste your own hand-written recap and just use Parcane as a campaign timeline and storage tool.
• Ask us anything — Have questions about how your data is handled? Email [email protected] and we'll answer honestly.