Skip to main content
The chat panel over the map: a question about foundational transformer papers, four Paper info tool calls, and the agent's answer

A turn: the question, the tools the agent called, the answer with arXiv ids. The ring in the header is the context meter.

Press ⌘K, click Ask the agent in the rail, or flip the bar’s icon to chat. Ten messages are free without an account; signed in, there is no cap. Sessions persist and are the same sessions the CLI lists with nisa sessions — start in one place, continue in the other. Terminal chat

​
What the agent can do

Search the corpus (text, semantic, by author, similar papers), read papers page by page, follow citation graphs, open a paper’s GitHub repository, search the web, run code in a CPU sandbox (numpy, pandas, scipy, matplotlib, scikit-learn preinstalled), draw charts and images, write PDFs, and manage your lists. Papers it cites are pushed onto the map and highlighted; in fullscreen (Esc leaves) a research rail shows the conversation’s papers with filters and CSV export.

​
Writing

Enter sends, Shift+Enter breaks a line. Send while it’s working and the message queues (× removes it). Stop freezes the answer; an error banner offers Retry. @ mentions a paper or one of your lists by name — the agent receives the id. / opens the commands: The small ring by the header is the context meter — hover for used / window; the agent compacts context itself when it fills, shown as Compacting context.

​
Research mode and sub-agents

For a broad question the agent may enter Research Mode: a card with the brief, a question list that fills in as Explorer runs answer each one, then a Research Report card with a download (a typeset PDF). Delegated work shows as cards — Explorer, Scout, Sandbox, Report Generator — click one to read what it did and download its report. Spend per research run is bounded.

​
Artifacts

Generated PDFs arrive as PDF ready cards with a download button. Generated images have a download button and open fullscreen on click.

​
Sharing a chat

/share invites someone by email — they need a RapidReview account; there is no public link. Everyone in the chat can contribute; messages carry the author’s avatar and the agent is told who said what. Your local-folder connection stays yours alone. Shared chats show a people icon in the rail. Deleting a chat gives you five seconds to undo.

​
Your local files

/connect shows a one-time brg_… token (five minutes). In a terminal, cd into the folder, run nisa, paste /connect <token>. The agent can now read that folder — writes and commands only if you allow them, and only from the CLI (/permissions). Local folder access Nisa is in preview — the Feedback item in the account menu goes straight to the team.