Private Chat
An ephemeral chat surface for conversations that should leave no trace. Nothing is written to disk, nothing is remembered between sessions, and the transcript is wiped the moment the overlay closes or Loach locks. Open it from the ghost icon in the title bar.
Nothing is persisted
Unlike regular chats, Private Chat does not create a chat row, message rows, metrics, or an attachment store. The transcript lives entirely in memory. The moment you close the overlay - or Loach locks - everything is gone, including the picked model, persona, tone, per-chat instructions, and parameters-panel state. There is nothing to export, nothing to search, and nothing for an attacker with disk access to recover. Private Chat never reads or writes Space or global memory either.
Ollama only
The model picker inside Private Chat only lists Ollama models, and starts on the Ollama model of your most recently active chat (or the first model in the list that can chat - embedding models are passed over). Ollama’s cloud models - tags ending in :cloud or -cloud, which Ollama runs on ollama.com - are left out. OpenAI-compatible providers are deliberately excluded - the data path for cloud providers crosses too many intermediaries (proxies, key vaults, hosted logging) to honour the “leaves no trace” promise.
Messages go to the Ollama server set in Settings → Providers. When that’s this machine, the empty chat reads Only local Ollama models can be used here. When it’s another machine, it says so: Only Ollama models can be used here, and messages go to the Ollama server at followed by its address.
No tools
The model inside Private Chat gets no tools at all. Servers configured in Settings → MCP are not exposed: a tool call could side-channel the conversation out to a third party - that defeats the point. The built-in tools are left out too, even those you have switched on, because the overlay has nowhere to show a tool call, its result or a file one makes. And a Private Chat never gets a folder, so the file tools from Work in folders can’t read or write files on disk. For anything that needs a tool, an MCP server or a folder, use a regular chat. See MCP support.
Two other features that would send your words elsewhere stay out too: links in your message aren’t fetched by web fetch, and there is no voice dictation button, because dictation can send audio to an online speech service.
Code blocks in replies keep Copy and Export, but there is no code canvas: Open and Preview aren’t offered, because the canvas belongs to the main window and would outlive the overlay.
Opening and closing
Open Private Chat from the ghost icon in the title bar, or with the /private slash command. It opens as a dark-only overlay above whatever you were doing, with the cursor in its composer. The overlay can only be dismissed by the explicit X in its header - clicking outside, pressing Esc, or otherwise stray-clicking does not close it. The wipe is destructive, so we don’t want one accidental click to throw away the conversation.
While Private Chat is open the regular chat surface is paused: any in-flight regular generation is cancelled, queued messages wait, a memory extraction still running for a regular chat pauses when you send a message here, and the rest of the app is non-interactive underneath until you close the overlay. Tab can’t reach it, the title-bar search and sidebar toggle are disabled, and app shortcuts are off - except Ctrl + Shift + L (⌘ + Shift + L on macOS), which still locks the app if you have set a lock up. Locking - with that shortcut or by auto-lock - ends the Private Chat session exactly like closing it, so the conversation is gone when you unlock.
Attaching files
Attach files with the + button in the composer. The same 20 MB per-file cap and handling apply as in a regular chat (see Attachments): text is added to your message, and images go to the model, sized by the app-wide Image scaling setting in Settings → Features - Private Chat has no setting of its own. The + button only attaches files: Library sources can’t be added here. Dropping files onto the overlay isn’t supported - the cursor shows the drop won’t be accepted - and a dropped file never lands in the regular chat behind it.
Shaping layers
Private Chat reuses the same prompt-shaping layers as a regular chat - just with a few deliberate omissions. Set them in the parameters panel (the sliders icon in the header); the active persona and tone show as chips above the composer, where you can remove them:
- Persona - any of the personas.
- Tone - pick any tone, or leave it unset to fall back to the default tone from Settings → General.
- Per-chat instructions - a free-form textarea, layered persona → instructions → tone.
- No Custom instructions or Space context - your global Custom instructions aren’t applied, and Private Chat is never attached to a Space, so a Space’s instructions, files and memory don’t apply.
- No
{{USER_NAME}}substitution and no temporal preamble - those belong to persistent chats.
Slash commands aren’t available in the Private Chat composer.
Parameters panel
A trimmed-down version of the regular parameters panel: Thinking (for models that support it), Context Length, Low VRAM and a reset button, with no Advanced view. A note at the top says whether you are on the app defaults, the model’s Modelfile defaults or your own changes. The request carries exactly what the panel shows - app defaults, then your global Thinking default and the model’s Modelfile defaults, then your changes, with the global Low VRAM setting pinning that switch on. The panel closes and wipes along with the rest of the overlay - nothing about it survives the close.
One-off questions you don’t want sitting in your history, throwaway experiments with prompts, and anything you’d normally type into a private browsing window. For ongoing work, a regular chat (or a Space) is the right tool - you cannot scroll back to a Private Chat tomorrow.