PAIE · Personal AI Environment

Transparency

🔴

REC indicator
When Active mode is on, your conversation is being saved to your vault. You control it.

⚠️

Cloud models
If you use a cloud API (OpenAI, Anthropic, Gemini), your messages leave your machine and are subject to that provider's data policy. PAIE warns you before this happens.

📋

Open prompts
Every system prompt PAIE uses is visible and editable here. No hidden instructions.

⬡

Your data stays local
Your files, your database, your machine. PAIE does not transmit vault data to any external server unless you explicitly use a cloud model.

Powered By

PAIE doesn't reinvent solved problems — it connects proven open-source engines instead of hiding what's actually doing the work.

🦙

Ollama
Local model inference — nothing leaves your machine unless you pick a cloud model.

🐘

PostgreSQL
File lineage, tags, revision history — the source of truth for your vault's metadata.

🔺

ChromaDB
Semantic search over your ingested content.

🐍

Ouroboros
The watcher that auto-ingests your watched folders as you add and edit files.

Coming soon: Kavita (self-hosted, GPLv3) for in-app EPUB reading, and an open-source STT/TTS engine for voice.

System prompts PAIE sends to the model. Edit freely — changes take effect on the next request, no restart needed.

Loading prompts…

Folders Ouroboros watches and auto-ingests. Core folders are fixed; you can add or remove custom ones.

Custom folders

How this works

🎙️

The flow
Mic → Speech-to-text transcribes what you said → shown in the input box so you can see exactly what was heard → sent through the normal chat pipeline (same Focus scope, same everything) → reply streams in as text → Text-to-speech reads it back out loud. Nothing is hidden — every step below is a real, swappable engine, not a black box.

Speech-to-text (what hears you)

Bigger = more accurate, slower. Applies to your next recording.

Text-to-speech (what speaks back)

54 real voices across 9 languages, from PAIE's own local Kokoro engine — pick one and hit Preview to hear it before committing.

Bring your own key. Local (Ollama) always works with no key — these are optional, for when you want a specific cloud model. Keys are stored server-side on this machine only, never in your browser, and are verified against the real provider before being saved. Model names route as provider:model (e.g. openai:gpt-4o-mini) — anything without a known prefix keeps working exactly as it always has, unchanged.

Loading providers…