Repository navigation
Leaf Bot 0.7.0
Added
- Microsoft Teams (Settings → Chat apps, for those who look after the server): a single-tenant Azure Bot's app ID,
tenant ID and client secret connect it. People in that Microsoft 365 organization link their Teams account under
Profile with a code, then write to their bots in a one-to-one Teams chat and get the replies there; a step that asks
first comes as a card with buttons, answered as in the app. Teams sends messages to the server, so it needs a
public HTTPS address; each request must carry a token signed with the Bot Framework's published keys for this bot
and this Teams service, or it is refused before it is read. Files, group chats and channels stay in the app for
now. - Reaching the server from your other devices over Tailscale (Settings → Connection, for the owner, off by default):
turning it on runstailscale serveso the server gets an https address that only the devices on your Tailscale
network reach, without opening it to the internet; signing in is still needed there. Nothing runs until it is
turned on, Tailscale keeps the setting across restarts, and the server answers at that address only while it is
shared. - Drawing on a picture before sending it: a pen button on an attached PNG, JPEG or WebP opens it to draw on in a
few colours and to pin numbered notes on it. "Add the marked copy" attaches the picture with the marks drawn in,
beside the untouched original, and puts the pins' notes in the message as a numbered list, so the bot sees both
where and what. - Bots make slide decks: a bot writes the slides (a cover, section dividers, titles with points, a paragraph, pictures
from its folders, speaker notes) and gets a PowerPoint file in the conversation, which Keynote and Google Slides
open too and the person can edit. Light or dark, with an accent colour; it costs nothing and reaches nothing
outside the server. - Bots make short videos (1 to 15 seconds, landscape, portrait or square) with the person's saved xAI key, the same
way they make pictures: the video is made as a job the server waits on (up to ten minutes), kept in the
conversation's folder and sent as a file. One video a turn, since each is billed by the second; it asks first like
a picture does. OpenAI's video API was shut down in September 2026, so xAI is the only provider for now. - Link previews: holding the pointer on a web link in a conversation shows the page's title, description and site,
and a GitHub pull request or issue also shows its state (open, draft, merged, closed) as an icon beside the link
and its author. The server reads the page the first time the link is hovered (only its start, without its query
string, private addresses refused, kept for ten minutes), so the site sees the server's address and that moment;
on the desktop app that is the person's own connection. No picture from the page is loaded in the app. A link
that carries something shaped like a key, a bot's private token or a long code is not read at all, so a link a
bot wrote cannot send data out on a hover. GitHub is asked without a sign-in, so only public repositories are
previewed. - An emoji button next to the message box: every emoji in sections, the ones picked lately first, and a search by
English name or keyword; the pick goes where the caret was, and Shift-click adds more without closing. The data
comes with the app, so it works offline. Phones keep their own emoji keyboard. - The desktop app checks spelling as you type: a misspelled word is underlined, and a right-click on it offers what it
could be and "Add to dictionary". It is on by default and turned off with "Check spelling" in the same menu or in
the Edit menu. On a Mac it uses the system's checker; on Windows and Linux it checks the system's languages it has
a dictionary for, then Turkish and English, which Electron downloads from Google the first time. - A message written while the app cannot reach the server is no longer lost or handed back as a draft: it stays in
the conversation marked as waiting for the connection, and goes out on its own when the connection is back (the
browser coming online, the live connection reconnecting, or a retry every 15 seconds). A second try of a send that
did arrive counts once, and the message then says it was written while offline. - Bots can offer work they notice (Settings → General, off by default). After answering a person, the bot's model
looks over the conversation once more and may offer one next piece of work it could take on, on a card: "Do it"
sends the task as the person's own message, "Not now" puts it away. It never offers to send, post, buy, book,
delete or change settings; a conversation is looked at at most every six hours and a person gets at most three
offers a day; and each look is counted under Work offers in usage. - A page that hides text from people but writes it to an AI agent ("ignore your previous instructions", "if you are an
AI agent…") is now caught when a bot reads it: in its browser (text pushed off the page, see-through, a few
pixels tall, clipped or drawn in the background's colour, and comments) and in a page fetched without a browser
(what the markup hides). The bot is told the text is the page's, not an instruction, and to say it found it; the
conversation also shows a warning naming the site, once per site in a turn. Hidden text that does not speak to an
agent, such as a screen reader's "Skip to main content", says nothing. - On the desktop app (the server on the person's own computer), a bot that asks first now writes and edits files in
its own folders without a card, named by their full path. No model judges the step, so nothing waits on one;
commands, the group's shared folder, anything outside the bot's folders and "always ask" rules still ask, and a
step the person just turned down is not let through. A server of its own keeps its wider rule for the bot's own
machine, unchanged. - A routine whose run finds one of its connections signed out now waits for the sign-in: its next scheduled runs
are skipped with the reason instead of each stopping at the same card, and the routine list says it is waiting.
The person is told once, outside the routine's conversation too: on the phone, in Telegram or WhatsApp, and in a
direct message from the Slack app when they linked their Slack account. The first slot after they sign in again
runs as usual; webhook, mail and Run now runs still go ahead and show the card. - The phone app's Today page now shows expired notes, security alarms (with "Mark as seen"), steps waiting for the
person's second yes, open promises with the reminder time, and routine suggestions; the profile settings hold the
routine-suggestion switch and the usage settings the promise safety-net switch. - The phone app's memory screen shows when a note is true until, lets the person set, change or clear that day when
editing, and lists expired notes; the permission rules list shows a rule's second person, rules can be added for
one bot and with a second person's yes (offered when two or more people could give it), and a chat card waiting for
a second person says who said yes first and offers no "always allow". - The phone app's bot-reply menu offers "Undo this turn" under the web app's conditions, listing what will be put back
and, apart, what cannot be, and shows the note it leaves as separate lines; a reply that made a promise shows the
promise line with Cancel; the routine screen has "Rehearse", with the instructions as typed, a replay of one of the
last runs, and the result beside that run's own report. - A routine can be rehearsed before it runs for real. A new or changed routine runs unattended and the person only
learned what it does from what it already did, so "Rehearse" in the routine's dialog runs it once, with the
instruction as typed (saved or not) and, for a routine something else starts, with what started one of its last
five runs, shown beside what that run actually reported (that input is now kept, sealed and cut to 20,000
characters, as long as the run is; older runs cannot be replayed). The bot works for the person who set the
routine up, in a conversation of its own that is in no list and goes after 7 days. Steps that only read (pages,
searches, files, mail and WhatsApp history, a connected app's read-only tools, commands that read the bot's own
folders) are really done; every other step is not done, the bot is told so and to go on as if it had been, and it
is listed as "would do" in the words the approval card uses. No card is ever asked, whatever the level or the
rules, and the private line still stops the turn. The result shows the report, those steps in order, the reads
that ran, how long it took and what it cost, which counts in usage as a rehearsal of that routine. Nothing of the
routine changes (schedule, last run, carried report, failure marks), nothing is posted to the bot's main
conversation or its Slack channel, and no promise look or undo line follows. Someone who may change the bot can
rehearse instructions they typed; the saved routine is rehearsed only by the person it works for. - A turn can be taken back. A bot's step can be wrong, and the person had to find and undo each change by hand, so
every turn now keeps, for 30 days, what its file tools wrote or edited, the table rows, tickets, routines, skills
and bot settings it changed, and the notes it wrote or changed, each with how it was before (the words of a note it forgot are never kept). A bot's reply
gets "Undo this turn" for the person whose turn it was (admins cannot take someone else's turn back): a list shows
what will be put back, newest first, and apart from it what cannot be, the steps that went outside (mail, chat
apps, the browser, programs, requests), forgotten notes and files too large to keep (2 MB each, 20 MB a turn). Anything changed
again after the turn, by a later turn or by hand, is left as it is and reported, a routine, skill, bot setting or
bot note is put back only by someone who may change the bot, and a turn can be taken back once, when it is no longer
running. The result is posted in the conversation. - Routines are suggested for what a person keeps asking by hand. People ask a bot for the same thing again and again
("this week's sales summary" every Monday) without thinking of a routine, so once a day, for each person who has
not turned it off, their own messages to bots of the last 45 days are grouped by meaning (when an embedding service
is set up) or by shared words. A request made at least three times on three different days to the same bot, that no
routine of the bot already does and the person has not turned down, becomes one small-model request on the bot's
own account (its cost shows on the usage page) that writes a routine name and instructions in the person's
language, with a schedule when the requests kept one (the same weekday, most days or most working days, around the
same hour) and none otherwise. Today lists them ("You asked Ada for this 4 times, on 4 different days") with the
instructions editable and "Create routine" or "Not this"; only someone who may change the bot can create the
routine, anyone else is told to ask its owner, and a request turned down is never offered again. Rooms, team
conversations, routine runs, API calls, text handed over from outside and chat-app strangers are never read, and
each person sees only their own suggestions. Each person can turn it off in their profile. - Promises are kept. Bots say "I'll let you know tomorrow" and nothing made it happen, so the prompt now tells a bot
to set up a one-time routine in the same turn it promises something later. As a safety net, when a reply in a
person's own conversation or room sounds like such a promise and the turn set up no routine, a short list of Turkish
and English phrasings decides whether one small request on the bot's own account reads it. A promise with a time
gets a one-time routine that does it then and tells the person in the same conversation, and a line under the
reply ("Promise kept on the calendar: tomorrow 09:00 · Cancel") takes it back; a promise that waits for something
("when the price drops") is only listed, with no time made up. Bots also record promises the person made to
others with a newpromise_tracktool, from the chat or from the person's own sent WhatsApp messages and mails
they read in a routine, and remind the person on the due day at a time they pick (09:00 by default). Open
promises, overdue first, are on the Today page with "Done" and "Drop", which also remove a routine still waiting;
each person sees only their own. The safety net can be turned off on the usage page, where its cost shows beside
the routine pre-check's. - A step can need two people. An "always ask" rule (with or without its amount limit) can now ask for a second
person's yes, for things like a card filled above 1,000 or agit pushthat one person alone should not decide.
The first yes works as before, from the app or a chat app; the card then waits as "waiting for a second person"
and shows who said yes, and the step does not run. The second yes comes from someone else who may change the bot
(or the one person the rule names), who sees the step and who asked for it, with the first yes and when, in a
"Waiting for your second yes" list on the Today page, and is told alone, without the conversation, which stays
private. The same person cannot give both yeses, a "no" from either refuses it, "always allow" is never offered
on such a card and a bot at full access still waits. If the second person is removed or loses the bot while it
waits, the card is refused with a reason, and a rule nobody else could answer is not saved. The bot's list of
steps names both people. - A bot that is being steered by text it read (a page, a mail, a webhook, another bot) is now caught by two
trip-wires that nothing the person asks for would ever touch. Every bot's prompt carries a private line it is told
never to repeat; if that line shows up in a tool's input (as it is, percent-encoded, in hex or in base64) or in
its reply, the step is refused and the reply is cleaned. And in Settings → API keys a person can add a "trap key":
a made-up key with a made-up address that guards nothing and is shown to their bots like any other key; naming it
in a request, asking for a key of that name, or reaching its address trips the wire. A trip stops the turn and the
work it handed to other bots, moves the bot to "Ask" if it was freer, says in plain words in the conversation what
happened and what the turn had read before, tells the person, and stays on the Today page until they mark it seen. - A memory note can be true only until a date. A bot that saves a fact that will stop being true ("working from
home this week", "Ali is on leave until the 5th") gives the date, worked out from today's date and the person's
time zone, and the note carries it into the bot's prompt as "true until ...". Once the date passes the note
reaches no prompt and no bot's search, so the bot stops acting on it; it is not deleted but moves to an
"Expired notes" list in Settings → Memory, where it can be deleted or given a new date, and the Today page counts
them. A note's date can be set, changed or cleared when editing it. - A long conversation can be summarised while it is idle, before the size at which a Claude bot would summarise it
in the middle of the next message (Settings → Model). Once a conversation is at least 70% of that size and has been
quiet for half an hour, the server folds its session and marks the place in the chat; a message that comes
meanwhile waits for the fold. Off by default, as after a pause the model reads the conversation again without its
cache; the fold counts toward spending, and a paused bot or one past its limit is left alone. Bots on another
model provider fold as they run. - A conversation is named by the bot's model once its first answer is in: a few words on what it is about, in place
of the first message's opening words. A title the person gives is never changed, a bot that is paused or past its
spending limit is not asked, and it can be turned off (Settings → General). - Settings → About has a check of the server's data for those who look after it: work that looks under way but that
nothing runs, approval cards waiting on such work, folders of deleted bots, conversations and groups, sealed content
that does not open, and the free disk space. A safe repair ends the hanging work as a restart would, marks the cards
expired, and moves the stray folders intodoctor-trashin the data folder instead of deleting them. - A bot with a mail address can type the one-time sign-in code a site mails to it into that site's own field
(browser_email_code), once a person turns this on for its mailbox (Settings → Mail). Only unread mail from the
open page's own site, from the last ten minutes, gives a code, and each mail gives it once; a code that cannot be
told apart is not guessed. The person can forward code mails from their own mailbox to the bot's address. - A task taught by showing it in the browser keeps its steps with the skill, and can be done again without the model:
from the play button by the skill in the bot's panel, or by the bot itself withskill_replay, which it tries
first for a shown task. Each step finds what it acts on by the words the person saw. The replay stops at the first
step that no longer fits the page, or that needs a password, and says which; the person can hand the rest to the
bot, and the bot reads the steps left and finishes with its own browser tools. - A bot that hands work to another bot can send files with it (
bot_sendtakesfiles): the other bot gets its own
copies underreceived/<sender>/in its conversation folder and reads where they are. The bot answering can send
files back with its answer (reply_files), and the asking bot reads where the copies landed. - A daily spending limit for the whole server beside the monthly one (Settings → Usage), and daily and monthly limits
for each bot in its advanced settings, with what it spent today and this month. A turn that would go past one does
not start and says which limit stopped it and where to change it. Another bot cannot set a bot's limits. - A bot, a group or every bot at once can be paused from the sidebar's menus. A paused bot's running work stops and
says why; what the person writes to it waits and is answered once it is resumed; its routines are skipped with the
reason; and another bot that asks it for help is told it is paused. A pause mark shows on its avatar and a bar above
the message box resumes it. Only admins pause a group or every bot. - A bot that finds a better way to do a skill's task while working for someone who may not change that skill now
suggests the change to the skill's owner instead of changing it for every bot that has it. The suggestion carries
the bot's reason; Settings → Skills shows it as a comparison with the skill as it is now, to accept or reject, and
the Today page counts the ones waiting. A newer suggestion from the same bot for the same person replaces the one
still waiting. Someone who may only talk with a bot can suggest changes too, but still cannot add skills to it. - Skills count how often bots read them, and the skills list shows the count.
- Bots can take an object avatar (Bot settings → Object): a mug, a lightbulb, a sprout, a game controller, a
barbell, a speech bubble or a watering can with eyes, in the bot's colour. The desktop and web app draw it with
volume; while the window has the focus it breathes, blinks and looks at the pointer, and it shows what the bot is
doing: dots while it thinks, ribbons while it writes or reads a page, an exclamation mark when it needs someone.
Avatars off screen rest, and the system's reduced motion setting stops them. The phone app shows the plain plush
for these bots for now.
Changed
-
Yumak Bot is now Leaf Bot. The repository is
esmworks/leafbot, the contact addressinfo@leafbot.app, the
imagesghcr.io/esmworks/leafbot-serverand-computer, and the app IDcom.esmworks.leafboton the desktop and
phones. What an earlier installation set up keeps working:- The desktop app is a new app under the new ID, so an installed Yumak Bot does not update itself to it: install
Leaf Bot from the release. On first start it takes over Yumak Bot's bots, chats and settings (quit Yumak Bot
first; it says so if it is still open). A desktop app that was connected to a server elsewhere starts with its
own server instead, since the saved key was sealed for the old app; connect it to that server again. - The phone apps are new apps too. A server that sends iPhone notifications needs the new app ID under its push
settings; until then they reach the old app only. sudo yumakbot updateon a server installed withdocker/install.shcarries the settings over: the.envkeys
get their new names, and the project, database, network bridge and Tailscale names stay as they were, so the
same volumes, data and address are found. The command isleafbotfrom then on;yumakbotstays as a link.- A server set up by hand keeps running with its old
YUMAKBOT_*settings, which are read wherever the new
LEAFBOT_*name is not set. Before renaming the keys in its.env, addCOMPOSE_PROJECT_NAME=yumakbot,
POSTGRES_USER=yumakbot,POSTGRES_DB=yumakbotandLEAFBOT_BRIDGE=yumakbot0, or Docker starts it with new,
empty volumes. A reverse proxy may keep sendingX-Yumakbot-Proxy. - Signed-in browsers and phones stay signed in,
yumakbot://links still open the desktop app, settings saved in
the browser come along, and backups and bot files made with Yumak Bot can still be restored and imported. The
app's own tools are namedmcp__leafbot__*; permission rules, tool limits, routines and chat history are
rewritten to the new name, also in a restored older backup.
- The desktop app is a new app under the new ID, so an installed Yumak Bot does not update itself to it: install
-
Someone with no account here (a member of a Discord server, say) can no longer have a bot sign in to a site: the
vault's password, its two-step codes and codes mailed to the bot are not offered in their turns, since those tools
run without a card. -
A skill a bot saves for someone belongs to that person, who may then change or delete it like one they wrote.
Before, only the admins could.
Fixed
- A bot whose model the provider retired, or whose model name it does not know, goes on with the backup model when
one is set, and the chat says to pick another model for the bot. Before, each message ended with the provider's
error. - A one-time routine whose run failed for a passing reason (a spent usage limit, a busy or unreachable provider) is
tried again, as a repeating routine is. A scheduled run that a server stop cut short is tried again too, once and
a little later, when it had let through no step that changes something. - A WhatsApp message that first arrives undecryptable reaches the bot once WhatsApp gets it again from the sender's
phone. It was taken for a repeat of the empty stand-in and dropped.