Each chat lives in its own editor tab (like a document). The left activity bar is a launcher: a scrollable Recent list (default last 30 days via grok.launcherHistoryDays), status dots, rename/delete, and New session (Grok or Claude). Full history and search live in the chat tab’s history popover. You can keep several chats open and switch between them like any other tabs.

A tab uses the whole editor canvas by default — no centred column, no width setting. Cards and capability rows flow into more columns as the tab gets wider; the transcript, composer, and code blocks are always full-bleed. The one exception is agent reply text, which keeps a left-aligned ~95-character reading measure so long paragraphs stay legible without narrowing anything else.

Closing a tab stops that chat’s live process but keeps history on disk (unless it was an empty “new session” you never used). After a VS Code reload, tabs can restore; the visible tab reconnects first.

The small line above New session reads e.g. v2026.8.9 · 1.2B tokens. That is not your usage — it is what building Grokbit itself has cost, aggregated across every maintainer and every session, baked into the build and identical for everyone (hover for the exact number and the date it was measured). Your own session's context lives in the composer donut and the status-bar percentage.

A compact Session setup chip on the left of the top bar opens Agent (Grok / Claude Code), Model, Thinking depth, and Mode (Agent / Plan first / Auto accept). The same popover is available from the model chip in the composer toolbar. A brand-new tab has no history, so changing any of them there is free and invisible — nothing to restart, nothing to lose. Claude omits the Thinking row (no reasoning-effort axis).

Grokbit Actions is on the welcome canvas of every new session, and any time from the Grokbit Actions button in the top bar or "Browse Grokbit Actions…" beside the composer's + button.

By default it shows the bundled Grokbit workflow as bordered tiles — five skills that ship with the extension and work identically on Grok and Claude Code (see the capabilities table above) — plus User Workflows from the active backend (Grok: .grok/workflows/*.rhai; Claude: .claude/workflows/*.js). User Workflow tiles use readable Title Case names (e.g. Create Workflow) in the same cyber green as the suite. On Grok, Create Workflow opens a Workflow Builder (required name + scope toggle, goal, phase/agent canvas). Craft with AI stays on that screen, auto-sends a structured brief for the create-workflow skill, shows a working status while the agent runs, then loads the proposed workflow back into the builder for editing.

How the workflows work (full technical): roles, loops, caps, artifacts, and human gates for each step — docs/grokbit-workflows.md. In the extension, open a workflow tile’s Details control for the same guide in-panel.

Click a tile to seed the composer (e.g. /grokbit-plan ) — nothing is sent until you press Enter. Your own skills, agents, and the CLI’s slash commands still work via / autocomplete. Set grok.actionsScope to all if you want Actions itself to list discovered skills, agents, and commands as well as the workflow.

The workflow skills are installed into ~/.grok/skills and ~/.claude/skills when the extension starts, and refreshed when it updates. Because that is a per-user install, they also show up in Grok and Claude Code sessions you start outside VS Code. Set grok.skills.provision to off to manage your own skills instead, or grok.showCapabilities to false to hide Actions entirely.

Thin-client Studio surfaces (not a separate Office app, not five permanent sidebar tabs):

• Docs (chat top bar) — browse a capped list of business documents already in the workspace; open, reveal in the OS, attach as context, or seed a path into the composer.

When a turn produces a document path, a document card still offers Copy path / Open / Reveal. Generation still uses CLI skills (/docx, /pptx, /xlsx, …) and ordinary file tools — ask in plain language or use slash skills from autocomplete. This is not a built-in Office suite.

Your last non-plan mode (Agent or Auto accept) is remembered for new sessions. Plan first is always a deliberate choice for the task at hand.

When the agent proposes an edit, the diff is shown inside the chat card — green additions, red deletions, long unchanged stretches collapsed. Buttons use plain language:

• Allow this change — approve once
• Allow always — approve this kind of action for the rest of the session (when the CLI offers it; path/command-bound when extractable)
• Don't allow — reject

The file is written only after you approve. Past edits stay reviewable via view diff on the tool row. Diffs never open a separate editor tab (so focus stays on chat).

Sometimes the agent needs a choice from you (multiple options). An inline question card appears; pick an option (and Submit if there are several questions). After you answer, the card collapses to a green “✓ your choice” summary. On resume, past answers replay as read-only cards.

• The active editor can be included automatically as context (setting: grok.includeActiveFileByDefault).
• Drag files from the Explorer, right-click → Grokbit: Send File, press Alt+G, or use the + button to attach files. Paste a screenshot into the message box (thumbnail appears above the input). Vision-capable agents (e.g. Claude when the adapter advertises image input) receive the image on send; Grok Build currently gets a file path plus a short notice that it cannot view images.
• Explicit attachments are stronger than the ambient “currently open” file.
• Hold Shift while dragging to embed file contents as a fenced block instead of a path reference.

The clock icon (and the launcher) lists this project’s sessions, newest first, merged across Grok and Claude with a backend badge on each row:

• Click a row to resume (conversation, images, plans, and tools replay on the correct agent).
• Hover to rename or delete.
• Search filters by name across your whole history.
• The list loads 100 at a time and loads more as you scroll (stays fast with thousands of sessions).
• Clear all history removes every session for this project except ones that still have an open tab.

Unread green/red badges persist across reloads until you open that session.

A native status bar item (bottom of VS Code) mirrors the active chat: model, thinking depth, mode, context %, working spinner, and a bell count when any session needs you. Click it to jump back to chat. Claude omits the effort segment (no reasoning-effort axis).

Optional setting grok.notifyWhenWaiting (off by default): a passive notification when a background tab needs you — it never steals focus; you choose Go to session.

Type /imagine <description> (or /imagine-video <description>) and results render inline in the chat. Hover for Copy path / Open in VS Code. Subscription-only Grok features; they survive session resume. Editing a reference photo with /imagine is supported when the CLI offers it.

The microphone in the composer records speech via xAI Speech-to-Text. Words appear live as you talk. Say your send phrase (default “grok send”) to submit hands-free.

Needs: ffmpeg on PATH + a separate API key (grok.voiceApiKey or env). Cost: pay-as-you-go STT (~$0.10/hr batch, ~$0.20/hr streaming) — not covered by SuperGrok alone. Details: research/voice-input.md.

While the agent works, each turn's activity rolls into one compact carousel strip instead of a scrolling list: the strip shows the current action (Reading chat.js…, Running command…) with a step counter, and ‹ › flips back through earlier steps. When the turn finishes it collapses to a one-line summary — e.g. Explored 5 items, edited 2 files · 9 steps — that expands to the full detail (tool rows, step narration, thinking). Commands can show output, failed tools turn red with a short reason, and a files changed strip above the composer summarizes applied edits this turn. Prefer the classic scrolling stream? Turn off grok.compactActivity (gear → Config & debug → Compact activity view).

Pick the model and reasoning effort (none → xhigh) from the Session setup chip (top bar), the model chip in the composer toolbar, the gear menu, or the command palette. Effort is how deeply Grok “thinks” (more depth = more tokens/time) and is per tab, not global. Changing effort restarts that session only (optional summarize-and-restart). Some model switches restart when the CLI requires a different agent type — the extension handles that. Claude has no effort axis; the UI omits it cleanly.

• grok.showThinking (default off) — when off, a short Thinking… stand-in is shown instead of full reasoning traces. Toggle live under gear → Config & debug → Show thinking details.
• grok.chatFontScale — zoom only the Grok chat (60–300%), not the whole editor.

• LaTeX math renders via MathJax (offline). Hover a display equation to copy or export PNG/SVG.
• Mermaid diagrams render inline (offline, theme-aware). Hover to copy source or export.
• Normal markdown (code, lists, tables) is fully supported.

Type / in the composer for autocomplete from your installed CLI (e.g. /imagine, /compact, …). Snapshot: docs/SLASH-COMMANDS.md.

The context donut shows how full the conversation window is. Use /compact (gear → Compact) to compress history, or start a new session. Lower thinking depth for cheaper/faster turns.

Extra tools (browser automation, etc.) are configured in the CLI (~/.grok/config.toml or project .grok/config.toml). Gear → Connected tools (MCP) / open config files, then start a new session to reload.

Gear → Log out of Grok/Claude (label follows the tab’s backend), or command Grokbit: Log Out — signs out that backend only, closes its live session tabs, and shows the signed-out launcher. History stays on disk; sign in again to resume.

Anonymous usage only: one session_start per real conversation with install id + mode/model/effort — never message content, code, or paths. Off with grok.telemetry.enabled: false or VS Code’s global telemetry setting. See docs/privacy.md.

Also listed in-app: gear → Keyboard shortcuts.

npm install
npm test         # grok-free unit/DOM/integration suite — same as CI
npm run package  # → grokbit-<version>.vsix

npm test never spawns the real grok binary. A separate npm run test:live exercises the real CLI before releases. See TESTS.md and docs/architecture.md.