Skip to content
VibeGuild
Menu
Sign inGet started

The side panel

The panel on the right is per-workspace and has five modes. Toggle it with v p in command mode, or jump straight to a mode: v e editor, v g git, v b browser, v k knowledge, v m MCP.

None of them are things your agents need — the agents have the terminal. They exist so you can follow along without leaving the app.

Editor

A Monaco editor (the one VS Code uses) with a file tree, preview tabs, and Cmd/Ctrl+S to save. It follows your app theme, down to syntax colors, and Settings → Editor owns its behavior.

The tree is a real file manager: right-click for a context menu, create and rename inline, duplicate, drag to move, and delete to the Trash rather than into nothing. Hold Option (Windows: Ctrl) while dragging inside the tree to copy instead of move.

Drag-and-drop is how the editor talks to everything else:

  • Drag a file onto a pane to insert a reference to it — @file for agents, the plain path for shells.
  • Drag a file out of the window to copy it into Finder or Explorer, and drag files in to copy them into the folder you drop them on.
  • Drop a file on the editor or its tab strip to open it; hold Shift to insert its path at the cursor instead.

Changed lines are marked in the gutter against HEAD, images preview inline, markdown files get a Code/Preview toggle, and a Files/Search toggle swaps the tree for fuzzy quick-open and full-text search (Cmd/Ctrl+P while the editor has focus). A selector above the tree reroots it into any of the workspace's live worktrees, with tabs kept separate per worktree — so reviewing what the agent on fix-auth did never disturbs the files you had open.

Git review

The git panel is branches first, because in this app a branch is usually an agent. Each row shows who's working there (the same numbered status circle as the pane — click it to focus that pane), whether the tree is dirty, how many commits it is ahead and behind, how much disk its worktree uses, and, when the gh CLI is installed and authenticated, its PR number and CI state.

Open a branch and you get two tabs:

  • Changes — what the branch would open as a PR right now, uncommitted work included, as side-by-side read-only diffs.
  • Commits — the branch's own log.

Row actions cover the lifecycle: open a pane on this branch (a fresh worktree runs the setup command first), reveal in Finder or Explorer, pin a worktree to exempt it from auto-cleanup, and delete. Branches with no checkout offer Delete branch — one click when merged, an armed confirm when there are unmerged commits — and a checkout row can delete the worktree and the branch in one gesture. Dirty deletions always ask.

"Merged" is smarter than the local graph: a branch whose PR GitHub reports as merged, with your local tip exactly at the merged commit, deletes in one click even though a squash or rebase left it looking "ahead". A Clean up button sweeps every provably merged, unoccupied, unpinned branch and its clean worktree in one confirmed pass — across every repo in a multi-repo workspace.

Checkouts other tools made (an agent's own sandbox, a manual git worktree add) show up too, badged external worktree: removable only when clean, never forced, and a dirty one can't be deleted here at all.

This is deliberately not a git client. Committing and merging stay in the agents' hands; the panel is for seeing what they did.

Browser

A real embedded browser — native webviews with real tabs, so background tabs keep running and popups open as tabs. The address bar searches or opens any http(s) page and autocompletes from your visit history (arrow keys, Enter; remembered across restarts, with a toggle and a clear button in Settings → Browser).

It is built for looking at what your agents built: back/forward, reload (Cmd/Ctrl+R), new and close tab (Cmd/Ctrl+T / Cmd/Ctrl+W), a devtools button that opens the Web Inspector, and downloads that land in Downloads with a notice.

The right-click menu is the interesting part. Besides copy, paste, and link actions, send to agent takes selected text — or a page element you pick visually — and drops it into the focused agent pane as pasted input with a trailing newline. It is never submitted for you. Agents can drive this from their side too: they may open localhost previews here and read back the console and the rendered page, which is what lets an agent look at its own dev server and fix what it sees.

One honest limit: while you're typing inside a web page, the page owns the keyboard. The app's shortcuts and the browser's own toolbar pause until you click any app surface — a pane, the toolbar, the rail — to bring them back.

Knowledge and MCP

The last two modes have their own pages:

  • Knowledge (v k) browses and edits what your agents remember and the skills they follow, including the exact instruction block every agent receives on connect. See Memory & skills.
  • MCP (v m) installs third-party MCP servers and decides which of your agents get them. See MCP servers.