Privacy & Security
Exactly what leaves your machine and what never does, feature by feature — plus how to run Triplo with no data leaving at all.
How Triplo handles your data
Three things shape everything below.
- Your content lives on your disk. Prompts, instructions, automations, conversations and knowledge bases are plain files in your Documents folder. There is no Triplo account holding them.
- Triplo is not a model provider. When you send a prompt it goes to the endpoint you chose. Choose a local model and it goes nowhere.
- Knowledge search is local. Indexing and retrieval run on your machine with a bundled model. Your documents are not uploaded to be searched.
What leaves your machine, feature by feature
| Action | Goes where | Carrying what |
|---|---|---|
| Sending a prompt | The model endpoint you selected | The full assembled prompt — instruction, Mind passages, conversation, your text |
| Sending a prompt to a local model | Nowhere | — |
| Searching a Mind | Nowhere | Runs entirely on your machine, against an index stored on your own disk |
| Enriching a new document | Triplo's endpoint | The document's text, once per document |
| Attaching an image | Triplo's vision endpoint | The image |
| Attaching or recording audio | Triplo's transcription endpoint | The audio |
| Scraping a web page | The website | An ordinary page request from your machine |
| Scraping a YouTube link | Triplo's transcript service | The video ID and your output language, never the prompt. See Scraping |
| Running an automation | The endpoint you configured | Your parameters and headers |
| Validating your licence | Triplo's licence service | Key, device fingerprint, CPU model, username, platform |
| Listing allowance models | Triplo's model service | Nothing about your content |
| Device sharing (if enabled) | Triplo's sharing service | Prompts, instructions, automations, Mind definitions |
| Checking for updates | GitHub Releases | Your current version |
| Saving a conversation | Nowhere | Local file only |
| Using the MCP server | Nowhere | Loopback only, never leaves the machine |
| Prompting Pi | The provider of the model the session runs on | The full prompt — and Pi then acts on this machine, inside the folder you approved |
| Prompting a Hermes gateway | The gateway you connected to | The full prompt — and the agent may then act on it with tools, on the machine it runs on |
The three that surprise people
Running local-only
It is possible to use Triplo with no content leaving your machine at all. It takes deliberate configuration.
Ollama or LM Studio as a Connection. Switch off Allowance Models so a hosted one cannot be selected by accident.
Pin the local model to every prompt that touches sensitive material, so the choice does not depend on what is selected in the top bar.
Search is local, but enrichment is not. Either accept one call per document, or add documents while disconnected and leave them unenriched — they stay searchable.
No image or audio attachments. Turn scraping off so a pasted link is not fetched.
If your licence has it, sharing uploads prompts, instructions and automations.
What still needs the network: licence validation at startup, and update checks. Neither carries your content.
The embedding model that powers Mind search is bundled inside the application — roughly 52 MB, never downloaded at runtime. A Mind works on a machine that has never had internet access.
Where credentials are stored
| Credential | Stored in | Notes |
|---|---|---|
| Licence key | config.json | Also included in a backup and in Copy Debug Data |
| Connection API keys | config.json | Never shared between devices |
| Automation headers | The automation's .md file | Plain text. Shared to other devices if sharing is on |
| MCP token | config.json | Generated once, reused |
None of these are encrypted at rest — they rely on your operating system's file permissions. Anyone who can read your home directory can read them. Use scoped tokens with the narrowest permissions that work, and rotate them if a machine is lost.
Three things that carry secrets
- Backups contain your licence key and every automation header. Store them like a password file.
- Copy Debug Data includes your licence key and an encoded copy of your whole configuration. Only send it to support.
- Shared automations carry their headers to every device that receives them.
Clipboard and selection
Two features touch things people reasonably consider private. Both are bounded.
//clipboard// is sent, or to show a short preview while that suggestion is visible in the list. There is no background monitoring. Turn it off with Listen to Clipboard.Captured only when you send a prompt containing //selected//, or when you press a SmartPrompt hotkey. Capture works by briefly synthesising a copy — and Triplo restores your previous clipboard afterwards, every time, including when capture fails.
Triplo deliberately never previews your selection, because previewing would mean copying, which would clobber your clipboard just to render a hint.
Application security
Always download from the Downloads page or the official GitHub releases. Builds from anywhere else are not ours.
Removing your data
Settings → Revoke. Releases the device registration.
Remove ~/Documents/Triplo AI/. That is every prompt, instruction, automation, conversation and Mind.
Remove the user-data directory listed in File locations. That is config.json, the logs, and Vectors/ — the vector indexes built from your Minds, which sit here rather than inside the Mind folders.
Drag the app to the Bin, run the Windows uninstaller, or delete the AppImage.
Nothing is left behind elsewhere on the machine. Anything you uploaded through device sharing is removed from the dashboard at control.triplo.ai.