You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Rooms keep meeting minutes. When a talk starts after a quiet spell, the room asks on a card whether it is a meeting
(yes, no, always in this room, never ask here); a meeting can also be started from the room's menu. It ends from the
header or after 30 quiet minutes, and the minutes (summary, decisions, tasks with who took them on and by when, open
questions) are written with one answer of a member's model, a long meeting in parts. The person reads, fixes and
approves them, and picks which bots are to know: each gets a note in memory with the decisions and the tasks it took
on first, so it knows them in every conversation, not only in the room. A note for a bot the person may only talk
with waits for the bot's owner. Minutes can be copied or downloaded as Markdown, and are kept encrypted.
Tasks in approved minutes that have a day can be reminded of. On that day, at a time the person picks, the bot that
took the task on looks at where it stands and tells the person, in the conversation where they talk; a person's
task is brought up by the first bot of the room the person may change. Each reminder is a one-time routine among
the bot's routines, and reminders that already ran are cleared away as new ones are set.
Settings → Import and backup reads a Grok Bot share link (x.ai/bot/…): it shows the bot's name, who shared it
and its description, opens the page to add the bot in Grok Bot and copies export instructions for that one bot.
The share page carries no more than that, so the bot itself still arrives in the import file.
The Grok Bot export instructions carry each bot's skills with their full text, its shape, and its integrations as
notes to set up again. The import preview shows each new skill's text and says which ones stay off because their
text hides something.
Bots on API keys run shell commands in the same OS sandbox as Claude bots: kept to the bot's folders, with an
environment of their own, the internet only when the bot has it on, and public addresses only. A bot with a
desktop keeps running its commands there.
Bots on API keys read web pages (main text, headings and links) and, when the server's environment names a
SearXNG with WEB_SEARCH_URL, search the web. Both ask under the same names as a Claude bot's steps, so the
same "always allow" and "always ask" rules and the same cards apply.
Opening a page whose address imitates a site the person uses (look-alike letters such as "faceb00k", one letter
off, or the name inside another domain such as "facebook.com.giris.io") asks at every level, full access too,
and the card warns which site it looks like. Such a card offers no "always allow". The sites measured against are
often copied ones (mail, social media, stores, payments, banks, e-Devlet), the sites the person keeps a sign-in
for, and the domains they wrote.
A routine can read WhatsApp: the chats people start with the linked account, and groups the person picks. A chat's
new messages reach the routine as outside text a couple of minutes after it goes quiet, with its last messages for
context, each chat in a conversation of its own. Messages are kept encrypted for 30 days; codes, passwords and
sign-in links are never kept, only a line saying one was left out. A bot drafts a reply or a WhatsApp Status post
on a card, and nothing is sent until the person approves it. Groups are only read, and only the person who linked
WhatsApp can have a routine read it. A bot can also set a routine to follow a group by its name or read people's
chats when asked in a conversation; it asks the person on a card first, at every level.
Changed
Bots work by themselves by default ("Full access"). A new bot, one a manager creates and one imported without a
level start there; the person can still pick "Ask" or "Read only". Only these still stop for the person, at every
level: their "always ask" rules, a page that imitates a site they use, filling a saved card, a sign-in, and the
commands no level lets through.
The "review" level, a small model judging each step, is retired. A bot still saved with it works at full access and
shows as such. It handed the person more cards than it spared them, and every step waited on a model call.
Text the bot read from someone else (a page, a mail, a webhook, another bot) no longer lowers its level. That text
still reaches the bot marked as outside data, and the steps above still ask.
"Always allow" on a card puts the bot on full access and lets through the cards it is waiting on, instead of
keeping a narrow rule for that one step. The suggestion for a bot whose cards the person nearly always allows now
offers full access.
A bot's memory changes take effect at once, wherever they come from: a routine, another bot, a turn that read
outside text, the background memory review, the shared memory, a place's notes, a pinned note or one the person
wrote. A stopped or failed turn keeps what it wrote. Every change still shows in the memory screen and its history,
where the person can change or delete it. Only a change to a bot's notes asked by someone who may not change that
bot still waits for someone who may; the background review skips such changes, and leaves the notes the person
wrote or pinned alone, instead of queuing them.
Fixed
The routine pre-check runs on the bot's own account, as its run does. It used to be Claude alone, so a routine of
a bot on API keys ran unchecked on a server with no Claude sign-in.
A command that only reads the bot's own folders goes without a card on its desktop too, and as a whole line:
readers chained with ; or &&, piped into readers, with errors thrown away and folders named in variables set on
the line (B=…; ls $B | grep …). Such a line opened a card. The bot's
other conversations and its desktop's home still ask.
The desktop app connects again as the bots' way out to the web after the server restarts. A restart used to end
the connection the way a takeover by another computer does, and the app stayed away until it was opened again.