Session Artifacts

A session produces more than conversation text. Its output falls into several categories, each with its own place:

Artifact list

The "Artifacts" section of the "Overview" panel lists the files this session changed or produced — web pages, images, documents, and diagrams. Open one to preview it in place, switch between source and preview, open it in a separate window, or reveal it in Finder. To browse the whole project directory and edit files directly, use the separate "Files" panel; see Workspaces, Git, and Changes.

Previews cover this session's working directory (every root of a multi-root workspace) plus the cache of files you uploaded. Files this session wrote or read, and images explicitly published from the managed output directory, can also be previewed. A standalone image embed in an assistant reply registers a script's saved screenshot, and reopening the chat restores that registration. The output directory supports legacy flat images; nested directories are restricted to their owning conversation. Other paths get an explicit "outside preview scope" answer. Strings that happen to appear in command output do not grant image access.

Artifacts render outside the turn fold: even if you collapse the process of this turn, the images and pages it produced stay visible. If an image file is cleaned up later, the card shows a load-failure notice instead of painting the filename as a broken image.

Files, websites, and media in replies

• File references are clickable: local files mentioned in a reply appear as file chips with a type icon, optionally with a line number; click one to open it in the right-side panel at that line.
• Website cards: http(s) links in a reply produce website cards below it. "Open with" picks where to open: in Klaus (the in-app browser panel), an installed Google Chrome, or the default browser; you can also copy the link. Each open uses one target, and the one you chose becomes the default next time.
• Inline audio and video: generated or referenced audio and video files get a player right in the conversation, with play, pause, and "Save a copy…", so you do not have to find the file first.

Pages and diagrams rendered in place

Klaus can render a page directly in the conversation for you to see, without you opening a file:

• Inline visualizations: it writes a self-contained HTML page into this session's dedicated visualization directory, then writes a one-line directive in its reply, and the transcript shows the rendered page in place. Only pages in that directory are rendered, and the content runs in a sandboxed iframe that cannot reach your other files. Each page is limited to 1 MB, and a message can show at most 8. This capability has a switch; turning it off makes the directory, the instructions, and the rendering all disappear together.
• Diagrams and 3D scenes: with official plugins such as Diagrams or 3D Studio installed, architecture diagrams, flowcharts, and 3D scenes appear as cards in the conversation and open in a panel for in-place editing.
• Interactive widgets: show_widget mounts an interactive HTML component in the conversation. It can be replaced wholesale or patched in part, and carries a title and a payload of data passed to the page. It is sandboxed as well, and cannot send a message on your behalf — sending requires your confirmation.

Run previews

run_preview opens an already-running http(s) preview address (for example a dev server) in a background browser tab, without pulling it in front of you. It only opens the page; it does not start the server. The address must be an absolute http/https URL, and URLs carrying credentials are rejected.

UI verification screenshots

After Klaus changes an interface and verifies it in the real running app, it can put the verification screenshots directly into its final reply so you see the result. Screenshots keep the original image and can be zoomed, saved, and found again in the history.

Images, video, speech, and PDF

Generating images, videos, and speech is a separate path: those three tools are always present, and whether they work is decided at each call from your choices under "Default models" — see Image, Video, and Speech Generation. Generated files land in the artifact list the same way.

PDFs are handled by the built-in PDF skill: extracting text and tables, rendering pages to check layout and charts, creating new documents, and splitting, merging, or rotating. The Python environment it needs is prepared automatically by Klaus the first time it is used, with nothing for you to install. It does not do OCR, so text in scanned documents cannot be read; PDFs in remote (SSH) projects have to be processed with the remote machine's own environment.