Help & Documentation

The sidebar reads the active tab — articles, documentation, or PDFs — and grounds its answers using local text extraction. Click the Page Context Chip above the chat box to toggle page reading on or off.

Drag and drop PDFs, Excel workbooks, CSVs, or source code files into the chat dock. 49 file types are indexed locally for Q&A alongside the page. Totals, averages and counts over a spreadsheet are calculated by Nightjar, not guessed by the model.

When using reasoning-capable models (e.g. Gemma 4 or Qwen), toggle 🧠 Reasoning to view step-by-step thinking traces before the answer.

Extracts repeated items on a listing page (products, articles, search results) into a structured table for CSV download.
Note: This is a beta feature. Enable it in Settings → Beta features.

For questions about a whole document: /find "phrase" your question. Searches the open page and attached files for the phrase, reads every place it appears (and any section headed risks, limitations and the like) in small passes, then answers with where each point came from, and says how much it read. Slower than a normal question, and far more complete: on a long article it found three times as many points. Give several phrases for the other names a document uses: /find "stratospheric aerosol" "SAI" what are its problems?

Reads the questions in a quiz or form on the open page, answers them with the loaded model and fills them in, reporting any it could not match. /quiz-check then marks answers already filled in.

Every request a page makes is checked inside the browser against EasyList and EasyPrivacy, using Brave's adblock-rust engine, before it is sent. Nothing is decrypted or proxied. Cookie consent banners are suppressed and auto-rejected.

By default, cookies and session storage are discarded when the browser closes. To stay signed in to a particular site, turn on the 🔑 Keep me signed in switch beside the address bar: that site stays signed in, and every other site is forgotten when Nightjar closes, as Chrome's and Firefox's clear-on-close settings do. To stay signed in everywhere, turn off Private browsing under Settings → General.

On a long page or file, Nightjar sends the model only the passages that match your question. Normally that is by words, so a question asked in other words can miss the passage that answers it ("how much output do they lose as they age" against a section on "degradation"). With Better search in long pages on, the EmbeddingGemma 2 model also finds passages by meaning. It runs on this computer and downloads once (310 MB). On a machine with a graphics card, Nightjar offers it the first time you ask about a long page; anywhere, it can be turned on in Settings → AI. Without a graphics card the first question on a long page takes several seconds while the page is read. Short pages are sent to the model whole and do not use it.

Hold 🎤 in the chat box while you speak, and let go when you are done: what you said is transcribed on this computer by the Whistle speech engine and put in the chat box to check before sending. Ten minutes at most. The 🎤 in the toolbar does the same for any text field on a page: click into an email, a message or a form field first, hold the 🎤, and what you say goes in at the cursor when you let go. Password fields are never written to; if there is nowhere to put it, the words are shown with a Copy button. Holding Ctrl+M (⇧⌘M on a Mac) works the same way; another shortcut can be chosen in Settings, under AI. When you stop, Tidy up shows what you said beside Tidy, where the local model fixes only spelling, punctuation and misheard words. Changes are marked, either side can be edited, and its Use this button puts it in place of your words; Cancel leaves them as they are, and Undo afterwards puts them back. English, German, French, Spanish, Italian, Dutch and Polish. The recording is never kept or sent. The first use downloads the engine (about 18 MB).

Use the password manager you already have. Nightjar Browser doesn't store passwords, so copy-and-paste and auto-type from 1Password, Bitwarden, KeePassXC or any other work as they do in any app; the window is titled after the page, which is how auto-type picks the right login. The assistant, the agent and the writing assistant never read password, one-time-code or payment-card fields.