SSH hosts in the agents view · Velork on the host · Sign in, the Velork way
Work on SSH hosts from any window, with a chat per host and Velork's own sign-in. Terminal permissions are rebuilt, slates get charts and inputs, and tabs, menus, notifications and tooltips look like the rest of Velork.
New
- SSH hosts in the agents view: start a chat on any SSH host from + or the folder menu, with no folder or remote window open. Each chat works on its own host, and one window can have chats on several hosts at once
- Velork on the host: the first connection installs Velork's server on the host (macOS and Linux, Intel and ARM) and keeps it running while chats use it
- Sign in, the Velork way: passwords, key passphrases and the host's questions are asked in one connect card that shows the progress, says when a password was refused, and explains in plain words why a connection failed, with Retry and the log one click away
- Pick a folder on the host: a folder browser lists the projects it finds on the host and common places, so you don't have to type a path (you still can)
- Everything on the host: the shell, file tools, search, Review, plans, designs with their images and videos, ⌘P files, file cards and links all work on the chat's host. Review runs git on the host and follows edits and commits there
- Add SSH hosts: add a host from the folder menu; recent hosts and the ones in your SSH config are listed with your folders
- Terminal permissions, rebuilt: one level per chat, rules that allow, ask or deny commands for a chat, a project or everywhere, and every part of a command checked on its own, so
git status; rm -rf ~no longer passes asgit status. Subagents follow their chat's level and ask in the chat you're in - Slates with charts and inputs: area, scatter, heatmap, histogram, funnel and treemap charts, inputs, markdown, code, key-value lists and timelines. Slates keep their state and can ask the AI, and slate calls show as plain tool rows
- Chat actions: right-click or double-click a chat in the sidebar, the session tabs or the IDE's chat tabs to rename it and more
Improvements
- Velork tabs, menus, notifications and tooltips: editor tabs get a + for a new file, an "N files" menu when they overflow and a hover card; notifications, menus and title bar tooltips use Velork's own look
- One model picker, one folder picker: the duplicate pickers are merged, Open Folder and Connect via SSH stay at the bottom, and long lists scroll
- Exact shell output: the AI reads exactly what a command printed (tabs, long lines and everything past the scrollback), not the terminal's screen
- Shell sandbox on macOS: a command can run so it only writes inside the project and temp folders, with no network unless allowed
- MCP servers that need sign-in: servers with OAuth (such as PostHog) sign in again, and keys are asked in Settings, Tools & MCP instead of a pop-up
- Mode suggestions: the "Switch to" chip follows what you ask for, not how long the message is
- Tokens follow the AI: while you follow Velork AI, the Tokens card opens and highlights the tokens it adds; otherwise a dot shows there are new ones
- Plans per chat: plans are saved in the chat's own project, and they work in windows with no folder
- Copying big designs: large copies on the canvas reach other windows
Fixes
- Host connections: a host that drops no longer breaks every other connection in the window, and host servers no longer restart every few minutes
- Restored terminals: terminals on a chat's host no longer come back on this computer with a "starting directory does not exist" error
- Cloud computer typing: paste works after a reconnect, and Ctrl no longer gets stuck down
- Review: Review keeps following changes after you reopen it, and diffs no longer show "couldn't be computed" after the file list changes
- Stop and queued messages: Stop ends the commands it handed off, force-send keeps the app servers you started running, and queued messages go to the chat they were written in