Slash commands
Commands start with / and are usually sent as a standalone message. They control the current session - switching models, adjusting reasoning depth, managing context, and more.
Commands only work for authorised senders - users on the allowlist or who have paired with the gateway. Unauthorised senders see /... tokens treated as plain text. If a command has no effect, ask your administrator to grant you access.
Commands vs directives
- Commands are standalone
/... messages (/reset, /status, /help).
- Directives (
/think, /verbose, /reasoning, /elevated, /exec, /model, /queue) can also ride in front of a normal message. A directive on its own line sticks for the session; a directive in front of a real message applies to that message only.
You can put an optional : between a command and its argument: /think: high, /send: on, /help:.
Session management
| Command | Effect |
|---|
/reset | Start a fresh conversation (new session) |
/new [model] | Same as /reset, optionally on a specific model |
/compact [instructions] | Summarise context and continue in a smaller window |
/stop | Stop the current running response |
/restart | Restart the agent (enabled by default) |
Model and reasoning
| Command | Effect |
|---|
/model <name> | Switch AI model for this session (/model alone shows a picker) |
/think <level> | Set reasoning depth: off | minimal | low | medium | high | xhigh | adaptive |
/verbose on|full|off | Show or hide detailed tool output |
/reasoning on|off|stream | Show the model’s reasoning in a separate message (stream = live draft on Telegram) |
/think notes:
- Send
/think with no argument to see the current level.
minimal thinks a little; high thinks the hardest; adaptive lets the model choose (the default on newer models).xhigh is only supported by certain models.
| Command | Effect |
|---|
/status | Gateway status, active model, provider quota |
/help | List available commands |
/commands | List available commands |
/whoami (alias /id) | Show your sender ID |
/context [list|detail] | Show how the context window is being used |
/usage tokens | Append a token count to each reply |
/usage cost | Print a local cost summary from session logs |
/usage full | Append a fuller usage footer |
/usage off | Turn the usage footer off |
Output and voice
| Command | Effect |
|---|
/tts always | Always send audio responses |
/tts off | Disable audio |
/tts tagged | Audio only for responses the model tags for voice |
On Discord the voice command is /voice, because Discord reserves /tts. Text /tts still works in a normal message.
Groups
| Command | Effect |
|---|
/activation mention | Bot replies only when mentioned or pinged (default) |
/activation always | Bot considers every message (and replies only when useful) |
/activation is owner-only and applies to the current group’s session.
Sub-agents and other sessions
| Command | Effect |
|---|
/subagents list | Show running sub-agents |
/kill <id|#|all> | Stop one or all running sub-agents |
/steer <id|#> <message> | Steer a running sub-agent (alias /tell) |
/queue <mode> | Control how rapid follow-up messages are handled |
| Command | Effect |
|---|
/exec host=<target> security=<level> | Show or set the tool execution policy (/exec alone shows current) |
/approve <id> allow-once|allow-always|deny | Resolve a pending tool approval |
/elevated on|off|ask|full | Control elevated (host) execution, when allowed |
! <command> and its alias /bash <command> run a shell command on the gateway host. They are disabled by default and require explicit administrator opt-in plus an allowlist. On most deployments they will do nothing.
Skills
| Command | Effect |
|---|
/skill <name> [input] | Run a named skill |
Inline shortcuts
A handful of commands work when embedded in a normal message and are stripped before the AI sees the rest:
/help, /commands, /status, /whoami (/id)
Example: Hey /status, and what's the weather in Sydney today?
Notes
- On Discord and Telegram, most commands also appear as native commands - type
/ in the chat to see them.
/verbose and /reasoning can expose internal tool output and reasoning - keep them off in shared group chats.- Command-only messages from authorised senders are handled immediately (they bypass the queue and the model) and bypass group mention gating.
