Commands come from two places: **shell builtins**, handled by the agent backend (xai-grok-shell), and **pager builtins**, handled by the TUI frontend (xai-grok-pager). Both show up in the same menu, and any enabled skill with `user-invocable: true` appears there too.
Every command below lists its aliases where it has them. A few commands only appear when a feature or session state enables them; those cases are called out inline.
Show how the context window is being used: a category breakdown (system prompt, messages, reasoning and overhead, free space) plus informational rows for tool definitions, the skills listing, and MCP server announcements with their estimated token cost.
Roll the conversation back to an earlier turn and discard everything after it.
### `/edit-prompt`
In minimal mode, open an external editor for an empty composer. Grok resolves `$VISUAL`, then `$EDITOR`, then `vi`; command values may include quoted arguments. Saving replaces the draft without sending it, and saving an empty file clears it. The command is hidden outside minimal mode.
To edit an **existing** draft when a terminal or multiplexer reserves `Ctrl+G`, open the command palette and select **Edit Prompt in External Editor**. That direct route preserves the existing text and refuses pasted, file-reference, or image chips without flattening them. Typing `/edit-prompt` into the composer necessarily replaces that input, so it starts from an empty draft.
Copy the most recent response to the clipboard. Pass a number to copy the Nth-latest response instead, or a file path to write the text to a file rather than the clipboard (handy over SSH, where the local clipboard is often unreachable).
Every copy is also written to a backup file — `~/.grok/last-copy.txt` by default, or `GROK_COPY_FILE` if set — and the toast tells you exactly where the text landed, so you can retrieve it even when the clipboard couldn't be reached or the copy went out as an OSC 52 escape this terminal couldn't confirm.
Switch models. Accepts a model ID or display name (case-insensitive), and for reasoning models you can add an effort level as a second argument. Alias: `/m`.
Set reasoning effort on the **current** model without reselecting it. Levels are `low`, `medium`, `high`, and `xhigh`, and it only applies when the active model supports reasoning effort.
Running one while the other is active switches modes — for example, `/auto` while always-approve is on switches to auto. `/auto` only appears when the auto permission-mode feature is enabled. You can also change mode with `Shift+Tab` (cycles Normal / Plan / Always-approve), `Ctrl+O`, or `/settings`.
Toggle multiline input. When it's on, `Enter` inserts a newline and `Shift+Enter` (or `Alt+Enter`) sends the message. Mid-turn, a bare `Enter` on an empty composer still force-sends the top queued follow-up. Alias: `/ml`.
For quick recall, press `↑` on an empty prompt instead. The panel opens with your most recent prompt already filled in; `↑`/`↓` step through entries (each lands in the input), `↓` past the newest entry closes the panel, and typing edits the recalled prompt in place.
Toggle vim-style scrollback keys (`j`/`k`, `h`/`l`, `g`/`G`, `y`/`Y`, and so on). With it off (the default), a bare letter or `Shift+letter` in the scrollback just focuses the prompt and types the character. The setting persists to `[ui] vim_mode`.
Reopen the current session in the other render mode. `/minimal` (offered while you're in fullscreen) switches to the experimental scrollback-native mode; `/fullscreen` (offered while you're in minimal; alias `/full`) switches back to the standard alt-screen TUI. Both relaunch the pager on the same conversation for this session only — they don't touch `config.toml`, and the relaunch banner reminds you how to switch back. The `--minimal` / `--fullscreen` CLI flags are session-scoped the same way. To make plain `grok` open in a given mode by default, use `/settings` → **Default screen mode** or set `[ui] screen_mode`.
`/flush`, `/dream`, and `/memory` require memory to be enabled (`--experimental-memory` or `GROK_MEMORY=1`); `/memory` also needs a configured memory backend. `/remember` is always available.
Save the current session's knowledge to memory right now, triggering an LLM summary of the most important content. Reach for it before compaction, or any time you want to lock in context.
Open the extensions modal on the Hooks tab, where you can view loaded hooks, add or remove custom ones, and toggle them individually. The modal does not grant project trust — see [10-hooks.md](10-hooks.md) for the trust model.
The shell also advertises individual `/hooks-list`, `/hooks-trust`, `/hooks-add`, `/hooks-remove`, and `/hooks-untrust` commands; in the TUI pager these are folded into the `/hooks` modal.
The shell additionally supports subcommands (`/plugins list`, `/plugins install <source>`, `/plugins uninstall <name>`, `/plugins update`, `/plugins reload`). In the TUI, the modal does the same work visually.
Intervals are `Ns` (seconds, minimum 60), `Nm` (minutes), `Nh` (hours), or `Nd` (days); anything under 60 seconds is raised to the minimum. Recurring tasks expire after 7 days, and you can cancel one with `scheduler_delete` using the job ID reported when the loop is created.
Set, manage, or check an autonomous goal. Grok works across rounds and only marks the goal complete after an independent evidence review confirms the claim; if that review can't reproduce the result or has no usable evidence, the goal stays active or pauses with concrete gaps.
Arguments are `<objective> [--budget <tokens>]`, or one of `status`, `pause`, `resume`, `clear`. The `--budget` here is a **token** budget for the goal run, separate from the agent-count budgets that workflows use. `/goal` appears when goal mode is enabled for the session. Which driver runs it depends on background workflows: with them on, the host evaluates each model round and runs adversarial verification on completion candidates; with them off, the legacy model-facing `update_goal` path reports progress and triggers verification.
Kick off a background research workflow. It plans a bounded set of questions, gathers structured claims with source evidence, cross-checks each claim on an independent verifier shard, and renders only the claims that survive, with their verified source locators. Failed shards, dropped claims, and researcher uncertainties are reported as coverage limitations, and the report is marked **Partial** whenever any remain.
```
/deep-research Compare the migration risks of PostgreSQL 17 and MySQL 9
The command returns right away — follow progress in `/workflows`, and the final report appears in the conversation on its own.
Model-launched workflows may set `agent_budget` on the `workflow` tool. It's an absolute cumulative cap on logical child-agent calls: every `agent()` call and every item in a `parallel()` panel spends one slot, while schema-correction retries don't. The default is 128, explicit values run 1–1,024, and a panel that would cross the remaining budget is rejected before any of its children launch. `budget()` reports the cap as `total`, admitted calls as `spent`, `reserved` (always zero), and `remaining`. Named slash launches use the default budget.
### `/workflow`
Launch a saved workflow, or manage a running one by the session-unique display name shown in `/workflows`. Launch the same workflow twice and the display names are numbered (`review-changes`, `review-changes-2`); you never need the internal run IDs.
Project workflows live in `.grok/workflows/*.rhai`; user workflows live in `~/.grok/workflows/*.rhai`. A same-process pause/resume continues the original immutable script, args, and `agent_budget` cap from committed host-call results — to iterate, edit the returned script copy and launch it as a new run.
A budget-limited run is different: it only resumes through a model/tool resume request that supplies an `agent_budget` above the admitted agent count. A bare `/workflow resume <name>` can't raise the cap, so it rejects budget-limited runs. Runs interrupted by a process restart aren't resumed at all, because external effects have no stable cross-process identity. And resume is not exactly-once: an external effect whose result wasn't committed before a same-process pause can run again.
### `/workflows`
Open the live workflows **run** dashboard — active and retained runs, not a catalog of saved definitions. Each row shows the run's display name, phase, agent roster, progress, and result. Inside a run's detail view, `p` pauses, `r` resumes an ordinary pause, and `x` stops. Budget-limited runs can't bare-resume: `r` returns the shell's rejection (raise the cap with a model/tool resume that passes a higher `agent_budget`), while `x` still stops. `s` saves the run's script, but it's hidden for known built-ins and numbered duplicate handles — for those, choose a new unique `meta.name` and save the edited script explicitly.
Send an aside to the agent without interrupting the current task. In minimal mode (`--minimal`), the answer shows up in a dismissible panel above the prompt: `Esc` dismisses it, a finished answer is saved into native scrollback, and a late reply to an already-dismissed panel is dropped. The side question and its answer aren't part of the main turn.
Check the current session for terminal, clipboard, color, input, notification, and sandbox issues. Doctor shows what it found and how to resolve each issue. Run `/doctor fix` to list available automatic fixes; other findings include manual steps. `/terminal-setup`, `/terminal-check`, and `/terminal-info` remain aliases.
Open the onboarding tutorial: a short list of topics (your first prompt, attaching context, navigation, slash commands, worktrees, plan mode, customization, switching from another agent tool) — each a ~30-second read, with `→` flowing straight to the next topic. Nothing auto-shows — this command (or the command palette) is the way in.
`/privacy` doesn't touch `[features] telemetry`, `trace_upload`, or your external OTEL settings — see [Monitoring Usage](24-monitoring-usage.md#related-settings). On team accounts, only a team admin can toggle privacy this way, and admins can also enable or disable Zero Data Retention for the team ([how to enable ZDR](https://docs.x.ai/developers/faq/security#how-to-enable-zdr)).
Any enabled skill with `user-invocable: true` in its SKILL.md frontmatter shows up as a slash command. (Turn a skill off via `/skills` and it stops being advertised.) So a skill at `~/.grok/skills/commit/SKILL.md` runs as:
Built-in commands always win over a skill with the same name. Name a skill "compact" and `/compact` still runs the built-in — but `/local:compact` invokes the skill.
The menu supports fuzzy search: start typing after `/` to filter. Each entry shows the command name, its description, an argument hint when it takes arguments, and its source (builtin, skill scope, or plugin name). Press `Tab` or `Enter` to accept the highlighted command.
