Sentient 3 is a desktop assistant that runs on your computer. It keeps your memory and tasks in one local file, works with a local model through Ollama or with the AI plans and keys you already have, and asks you before anything leaves your computer. This is the first beta of v3: try it, and tell us what breaks.

Highlights

Tasks that really run. Ask in plain words for a one-off job, a routine ("every morning at 8…") or a watch on a page. Sentient plans it, asks you to approve the plan, runs it, and checks what actually happened instead of trusting the model. You can try a task first, pause it, and see stuck or failed runs plainly.
Memory you can see and control. Sentient remembers what you tell it and keeps work, personal and project memories apart. It shows where each memory came from and how sure it is, and lets you correct or forget anything. Disconnecting an app can also forget everything learned from it. You can import your memory from ChatGPT, Claude or OpenClaw.
Your apps and devices. Gmail and other email, calendars, Home Assistant (devices are checked after every change), your notes and Obsidian vaults, MCP servers, a browser it can use for you, a terminal, and code it runs in a sandbox.
Talk to it anywhere. The desktop app, push-to-talk and dictation into any app, Ask on selected text, alarms and reminders, plus Telegram, Discord and WhatsApp, group chats included, where only you can steer it.
Safe by default. Anything that sends, buys, deletes or runs code asks first. A second look checks every recipient. Content written by someone else can't steer what Sentient sends. Undo a send, use Stop everything, and keep it all local with one switch.
Your models, your plans. A local model sized for your computer, any API key, Sign in with ChatGPT, a Claude Max or Team plan's API credits, or your own Claude Code for chats (experimental). Each chat can keep its own model.

Coming from Hermes? Sentient can import your setup.

Install: download the installer for your system below. Windows: Sentient-Setup-3.0.0-beta.1.exe. macOS: the .dmg for Apple silicon (arm64) or Intel (x64). Linux: the .AppImage. Uninstalling never deletes your data folder (~/.sentient).

What's new in detail

Added

Disconnecting can also mean forgetting. When you disconnect Gmail, IMAP, Home Assistant or another app, remove an
MCP server, a device or a folder of notes, or disconnect Telegram, Discord or WhatsApp, you can tick "Also forget what
Sentient learned from it". The dialog says how much that is first, then the memories, chats, summaries, saved emails
and app results that came from it are deleted, and Activity notes it. A memory you also shared somewhere else stays.
Leave the box clear and Sentient keeps what it learned (#153).
Skills can say how to check they work. A skill can list a file that must exist or a few lines of Python with the
output they should print. Sentient checks a skill it drafted for itself before it shows it to you, and doesn't
propose one that fails. A check that runs code waits until you approve the skill (the button says "Approve and run
check"), then runs in the code sandbox without your keys, and the skill is only saved if it passes. The Skills
screen shows each check and how it went (#386).
Home Assistant devices are checked after Sentient changes them. A moment after it turns on a light, locks a
door, opens the blinds, arms the alarm or sets the thermostat, Sentient asks Home Assistant for the device's state
again, for a few seconds, and only calls it done when the device really changed. If it didn't, you see it plainly
("Hallway light still reports off."), in chat, in quick commands and in task runs, and never a success. You can
change how long it waits in Settings (#386).
Folders and chat apps follow your memory places too. In Settings, Folders, each folder or Obsidian vault now has
a Used in choice: every chat (as before, and what your folders already have), Personal, Work or a project. A
folder kept for Work only comes up in work chats, so a personal chat never brings up your work notes unless you ask.
In Telegram, Discord and WhatsApp, send /scope work, /scope personal or /scope followed by a project's name to
pick which memories that chat uses, or /scope to see which it uses now. On the Memory page, where a memory is
kept now reads "Used in Work" next to its source, so it no longer looks like a topic such as "Work & Learning".
Say "stop" or "snooze" to a ringing alarm. While a Sentient alarm rings, typing or saying "stop", "stop the
alarm", "turn it off", "snooze" or "snooze for 5 minutes" in a chat, by voice or in Telegram, Discord or WhatsApp
stops or snoozes it at once, with no model ("Alarm stopped.", "Snoozed until 7:39 AM."). A snooze is 9 minutes
unless you say how long or change "Snooze length" in Settings. The banner has a Snooze button too, and after a
snooze it shows until when, with Turn off. When no alarm is ringing, these words go to the assistant as before, so
"turn off the alarm" can still mean your house alarm (#390).
Sentient in group chats, safely. You can now pair a Telegram group or a Discord server channel with the same
/pair code. Sentient answers only you there, when you mention it, reply to it or send a command. What other people
write is treated like an email from a stranger: Sentient can read it, but never takes orders from it, and after
someone else writes, anything it would send checks with you first, naming who wrote it ("That address came from
another group member (Ravi), not from you."). Only you can press its buttons, and task results stay out of a group
unless you turn them on.
Trust a chat again after you've checked it. A chat that read an email, a web page or a group member's message
shows a bar saying so above the message box. If you've looked at it and it's fine, click I've checked it (or
send /trust, then /trust yes, in Telegram, Discord or WhatsApp) and sending from that chat follows your usual approval settings
again. It covers only what you've seen: if someone writes while you decide, the chat keeps asking and says who
wrote. It's noted in the chat and in Activity, and anything new from outside makes it ask again.
Keep work, personal and project memories apart. Each memory is now kept for Personal, Work, a project you name,
or every chat (for things like your name or how you like answers). Pick a chat's memories from the button at the
top of the chat, or before you start one. A work chat then uses your work memories and the ones for every chat,
and leaves out your personal memories (like family or health) unless you ask about them; what Sentient learns
there is kept for work too. How you like things done (your answer style and other preferences it has picked up)
still applies in every chat. Tasks a chat creates keep its memories, and each messaging app can be set to one in
Settings. On the Memory page you can show one place at a time, move a memory to another place, and forget
everything kept for one place without touching the rest. Everything you had before is in Personal.
Sentient checks what really happened, not what the model says happened. After it writes a file, sends an
email or a message, creates a calendar event or runs a Home Assistant action, Sentient checks the result itself: the app said it worked, and
where it can, it reads it back (the file holds what was written, the email is in your Gmail Sent folder, the event
is on your calendar). MCP tools are checked too, including a send that worked for some people but not others. When
something didn't happen, a task run is marked Needs attention instead of done and says so plainly ("Sentient
said it sent the email, but mailer reported a problem: it couldn't do it for bad@example.com (mailbox not
found)."), and so does a chat reply. A clean run shows "Checked: file written, email sent."
Home Assistant, your smart home. Connect it with your Home Assistant address and a long-lived access token (kept
in your system keychain), then choose which devices Sentient may use. Sentient finds the devices it needs by room,
kind or name, turns lights and switches on and off, sets brightness and thermostats, and runs scenes and scripts.
Locks, alarms, garage doors and gates always ask you first, with a plain warning and only "Allow once"; unlocking,
turning off the alarm and opening a garage door or gate ask even when approvals are off. Tasks can
start when a device changes, like "when the front door unlocks". Quick commands now work for lights too: "turn off
the living room lights" happens at once, with no model, when it is clear which lights you mean (#175).
Alarms, "what's next on my calendar", and reminders on your phone. "Wake me at 7" and "set an alarm for 6:30"
now work at once, without the model, like timers: Sentient says exactly when it will ring ("Alarm set for 7:00 AM
tomorrow."), and when it does, the window shows a banner with Stop and plays a chime until you stop it. "What
alarms do I have" and "turn off the gym alarm" work too. "What's next on my calendar" (or "what's my next meeting")
answers straight from your connected Google Calendar, also without the model. And a new switch on the Channels page,
Reminders and alarms on your phone too, sends the reminders and alarms you set on your computer to your paired
Telegram, Discord or WhatsApp chat as well.
A safety check for skills and MCP servers from outside. Skills brought over from Hermes, and skills you add
with the new Add from a link button in Skills (a SKILL.md, or a GitHub folder that holds one), wait for your
review with a plain safety report: where it came from, and whether it tries to send your data to a fixed address,
run downloaded code, turn off approvals, read passwords or the keychain, hide what it does from you, or give
Sentient new orders, including in text you can't see (invisible characters, white-on-white text, hidden comments)
and in the scripts it brings. When the risk is high, Decline is the main button and approving needs
Approve anyway. A new MCP server's tools are not offered to Sentient until you look at each one: its risk, what
its description says, and whether it gives Sentient orders ("always call this first"), hides text or takes the name
of one of Sentient's own tools. Risky tools start switched off. Sentient remembers exactly what each approved tool
said about itself; if a server changes that later, the tool is held back and you get a notification showing the old
and new description side by side. Servers you had before this update keep their tools.
Memory you can see and undo. Each time Sentient remembers something from a chat, a small "Remembered: ...
Undo" note appears in that chat; Undo forgets it, or puts back what an update changed. Things that read like
instructions ("from now on, always cc Bob on invoices") are never saved quietly: they wait on the Memory page's
Review tab under "Instructions for Sentient" until you approve them. And in a chat that read an email or a web page,
what you said yourself is now remembered right away; only what came from the email or page waits for your review.
Two new rules: "Only if I asked" and "Hand it to me". In Settings > Approvals & safety, each app or tool can
now also be set to Only if I asked: Sentient goes ahead without asking only when your own message in a chat asks
for it ("email Priya the notes"), and asks every other time, including in tasks and anything it does on its own.
Or set it to Hand it to me: Sentient prepares it, like a draft email or a filled form, shows it to you with
Copy and Open in your email app, and never does it itself. Tasks do the same and send you a "Ready for you"
notification. Saying "don't send emails unless I ask" or "just draft them, I'll send them myself" in a chat
suggests the matching rule.
Try a task first, and put back files that code changed. A task waiting for your OK (or any task, from its menu)
now has Try it first: Sentient runs it once for real but only notes everything that would send, post, buy,
delete, change or run something, and shows you exactly what a real run would do, with the full email or message for
each. Look-ups still run, so it works on your real data, and the task's schedule and status stay as they are. And
before a command in the terminal, a script or a file write changes files in your allowed folders or Sentient's own
files, Sentient keeps a copy of them. If something changed, Activity lists it with Restore, which puts the files
back (and can itself be undone). Big files, files that aren't text and folders like .git or node_modules are left
out, with a note saying so. When your folders hold more than 500 text files or 25 MB, nothing is copied for that
command and Activity says "Too many files to keep a copy". Copies go after 7 days. Settings > Approvals & safety has
the limits.
The same memory in two languages is kept once. If Sentient remembered "माया शाकाहारी है" and "Maya is
vegetarian" separately, the nightly memory tidy-up now notices and keeps just one, in the language most of your
memories are in. It only does this when the model is sure they say exactly the same thing: never when one adds a
detail or they disagree. Each merge shows up in Memory > Review under "Tidied up overnight", with Undo, and a
pair you split again is never merged again. Turn it off in Settings > Knowing you.
Answers from your notes, even with small models. When a message clearly matches something in your folders or
Obsidian vault, the two or three best matching pieces now go along with it, so local models like qwen3:8b answer
"when does my insurance renew?" from your notes without having to search them. Only short pieces go in, so replies
stay quick on small models, and the notes a reply used show as chips under it. Turn it off or change how many pieces
go along in Settings > Folders.
Timers and reminders, instantly. "Set a timer for 10 minutes", "remind me at 6 pm to call Ravi", "remind me in
20 minutes", "what timers do I have" and "cancel the pasta timer" now happen at once, in the chat, by voice and
in Telegram, Discord and WhatsApp, without waiting for the model, with a short confirmation. When a timer or
reminder is due you get a notification at the right time in your timezone, and one set in a messaging app goes off
in that chat too. Activity lists the ones coming up, with Cancel, and the ones that went off. Anything that could
mean two things ("remind me at 6", "every day at 6") still goes to the model, and nothing quick ever sends, buys or
deletes. English for now; turn it off in Settings > General > Conversations ("Quick timers and reminders"). There are no smart-home
commands yet, since Sentient has no lights integration.
Undo a send, see exactly what goes out, and the trash by default. When Sentient sends an email, a message or a
post from a chat, it now waits 10 seconds first and shows what goes out, with Undo and Send now, in the chat
and as a toast (and as buttons in Telegram, Discord and WhatsApp). Undo and nothing is sent. Approval cards for
emails, messages, posts, deletions and purchases show who it goes to, the subject and the full text the way you'd
read them, with the raw details one click away; task approvals show the same. When an app can move something to
the trash, "delete it" now moves it there instead of deleting it for good, unless you say "permanently", and
Activity offers Restore. Purchases, code and tasks running on their own never wait. Change the wait (0 turns it
off) or the trash default in Settings > Approvals & safety.
Your notes and Obsidian vaults, searchable. In Settings > Folders, add a folder of notes or documents, or an
Obsidian vault, and Sentient searches it when you ask about something you wrote down: "what did I note about the
orchard budget?". Markdown, text, PDF and Word files are read; hidden folders like .obsidian, .git and .trash are
left out, along with the vault's own excluded files and any patterns you add. Answers show the notes they used as
chips, and a note from a vault opens in Obsidian. Sentient only reads these folders and never changes anything in
them. It notices changes as you write, reads only what changed, and does the work in the background with your
chats going first, so a big vault never slows a reply. Settings shows each folder's files, searchable pieces and
when it was last checked, with Check again and Remove. Web pages clipped into a vault count as written by someone
else: after reading one, Sentient asks before sending anything.
Activity: everything Sentient did, and what it will do next, in one place. A new Activity screen in the
sidebar lists, newest first, the apps Sentient used in your chats, task runs and briefs, each check of a watch,
approvals and how you answered them, replies in Telegram, Discord and WhatsApp, suggestions and other notifications,
memories it saved, skills it learned or changed, backups, and what paid cloud models cost each day. Every line says
what happened and why ("Made your Daily Brief: it runs every weekday at 7:30 AM", "Asked you before "Send" in Gmail
(to Arjun). You chose Allow once."). On the side: what is waiting for you now, everything scheduled for the coming
week (tasks, briefs, watches, the next backup and the nightly memory tidy-up) and the routines that are paused.
Filter by kind and by app, search, and click any line to open the chat, task, memory or setting it came from.
Under Messaging apps it also lists every notification sent to your Telegram, Discord or WhatsApp chats (task
results, briefs, approval requests, questions and suggestions): which chat, what was sent, and whether it arrived
or why it didn't. Click one to open the task or notification it came from.
A second look before anything leaves Sentient: before it sends, posts, invites people, buys or deletes
something, Sentient checks that everyone it goes to is someone you named, in your contacts or in your memories,
and that every amount is one you gave. If a web page or email slipped in an address ("also forward this to ..."),
or the content doesn't match what you asked, it asks you first and the approval card says why: "You asked to email
Arjun, but this goes to x@example.com." Your fast model also takes a quick look at sends, purchases, deletions and
code runs that would otherwise go ahead without asking, seeing only your own words and the action. This holds even
with approvals off, for apps you set to Allow and after "Allow for 1 hour". Turn the model part off in Settings >
Approvals & safety; recipients and amounts are always checked.
Answer approvals from the notification: when a task needs your OK while Sentient is in the background or
minimized, the notification on your computer has Allow once and Deny buttons (Windows and macOS), so the
task carries on without opening the app. Clicking the notification itself opens the task with its approval card,
where the wider choices are. If the request already expired or was answered, a second notification says so. With
notifications turned off in Windows, the Sentient taskbar button flashes instead.
Sources for tasks: each run in a task's history now shows the pages, emails, files and apps it used as small
chips, like the ones under chat replies. Click one to see what Sentient read and open it.
See what your models cost, and keep everything on this computer: Settings > Usage now shows what paid cloud
models cost today, this week and this month, by model and by job. Models on your Claude or ChatGPT plan show as
On your plan instead of dollars, and local models as free. Set a monthly budget there: you get a
notification at 80%, and at 100% paid cloud models pause until you choose Keep going this month (or, if you
prefer, until next month). The model menu in the title bar has a new Local only / May use cloud switch. In
Local only, nothing is sent to a model outside this computer: not a backup model, not a chat with its own cloud
model, not tasks, memory search or voice. Anything that would have used the cloud says so in plain words, or a
local model answers instead.
Ask about selected text, in any app: select text anywhere (an email, a web page, a document) and press
Ctrl+Alt+Shift+A (Command+Option+Shift+A on a Mac). A small window opens by your pointer and shows exactly what
Sentient copied, with Explain, Summarize, Translate and Reply, or ask your own question. The answer streams in, and
Copy or Replace selection puts it back where you were writing. Nothing is read until you press the
shortcut, your clipboard is put back right away, a password box is never copied, and nothing goes to Sentient
until you pick an action. The text counts as outside content, so Sentient never follows instructions in it and
asks before sending anything. Turn it off or change the shortcut in Settings > General.
Approve from your phone, and choose how far a yes goes: when a task needs your OK and you're away from the
app, the request goes to your paired Telegram, Discord or WhatsApp chat with buttons (or reply with once, task,
hour, always or no). Only the chat it was sent to can answer it, and a request that waits too long (4 hours by
default) stops the task instead of being approved late. Every approval now lets you pick Allow once, Allow
for this chat or Allow for this task, Allow for 1 hour, or Always allow (which saves an Allow rule).
Purchases, running code and anything after Sentient read an email or web page still ask every time. Settings >
Approvals & safety lists what you allowed for a task or for a while, with Revoke. Tasks that reach a tool you set to
Ask now pause and ask you instead of failing.
Routines pause while you're away: when you haven't used Sentient for a week (the app or a chat), or 20 results
have piled up since you were last around, Sentient asks once whether to keep your recurring tasks and briefs
running. If you don't come back within a day, it pauses them and tells you in one notification. Any sign of you,
in the app or a paired chat, keeps them running. Resume routines in that notification, in your chat app or on
the Tasks page brings them all back. Turn on Keep running while I'm away on a task that should never pause.
The day counts and the waiting time are in Settings > Tasks.
Sentient never overwrites a skill you edited: when you change a skill, in Skills or in any text editor, Sentient
notices and leaves your version alone. When it has a better idea it drafts the change for you to compare and
approve, and says plainly that your version is untouched. A proposal drafted before your edit warns that approving
would undo your changes. Reset to original puts back the version from before your edits.
Skill checks: Skills now point out problems in plain words: a missing or unreadable header, no description or
steps, a skill that is too long, links to files that aren't there, tools that aren't installed, a skill that reads
like a diary of past runs instead of a lesson, and risky instructions (skip approvals, ignore other instructions,
wipe folders, run a downloaded script, send passwords or keys). They show on each skill, on proposals, on skills
coming from Hermes, and a skill file that can't load at all is listed with what is wrong. A skill Sentient writes
with a risky instruction always waits for your review.
Bring your memory from ChatGPT, Claude or OpenClaw: Memory > Import > From another assistant reads a ChatGPT or
Claude data export (the zip they email you) or your OpenClaw workspace folder, right on your computer. Nothing is
uploaded and the export is never changed. You first see what's in it (saved memories, conversations, Claude's
memory and project instructions, OpenClaw's MEMORY.md, USER.md and daily notes) and pick what to bring over. Saved
memories come over as they are; for conversations and notes your fast AI model picks out lasting facts about you
in small batches. Everything waits in Memory > Review, and each card says which app, conversation and day it came
from. Things Sentient already knows are skipped, also across apps, and importing a newer export later only reads
what's new. You can stop an import, and remove everything from one app in one click.
Export your memory: Memory > ... > Export saves everything Sentient remembers about you as a readable Markdown
file or a SQLite database file in Sentient's files folder, grouped by topic and saying where each memory came
from. Nothing leaves your computer.
Each chat remembers its model: pick a model in a chat's message box and that chat keeps it, also after you
reopen it or restart Sentient. Other chats keep following your default model, and switching the default (or a
preset) in the title bar only changes those. Use the default puts a chat back on the default. The chat list
shows a small label on chats with their own model. If a chat's model can't answer (Claude Code is turned off or not
signed in, a local model was removed, a key is missing), the default model answers instead and the reply says so in
one plain line. Helpers a chat starts use its model too. In Telegram, Discord and WhatsApp, /model chat <model>
does the same for one conversation.
Backups and restore: Settings > Backups backs up everything Sentient keeps (chats, memory, tasks, settings,
skills and files) to one file, by hand or every day or week. Backups go to Sentient Backups in your Documents
folder (the one Windows really uses, also when it lives in OneDrive), apart from your data, or to a folder you
pick. Settings says when that folder syncs to OneDrive, Dropbox, Google Drive or iCloud, since the backups are
then copied to the cloud too. A backup holds the certificate your devices trust, so keep
backups somewhere private. Restore one from the list or from
a file: you see what is in it and confirm, Sentient keeps a snapshot of what you have now, then restarts with the
backup. Passwords, API keys and sign-ins stay in your computer's keychain and are never in a backup; after a restore
Sentient lists the ones to sign in to again. Sentient also takes a snapshot on its own before it updates its
database, before a Hermes import and before a restore, keeps the last five, and any of them can be restored.
Recurring tasks remember their last runs: each run of a recurring task now reads a short note of what its last
three runs found, sent or decided, plus the task's own notes, so it tells you only what is new instead of repeating
yesterday's news. Runs keep those notes up to date themselves; see, edit or clear them under Notes between runs
on the task.
Watches that only speak up when something important changed: a check script can now Tell me when something
important changes. Each check is compared with the last one without any AI, so an unchanged page costs nothing
and stays silent; only when something changed does your fast AI model decide whether it matters, and then you get
one notification that says what changed. Every watch also keeps a quiet list of its recent checks ("Checked,
nothing new", "Something changed. You were told") under Runs. When you ask Sentient to watch a page, a feed or the
news, it now uses this kind of check by default, so a view counter or a new timestamp no longer sends you a message.
Fuzzy times for repeating tasks: a daily or weekly task can run "in the morning", "afternoon", "evening" or "at
night" instead of at an exact time; Sentient picks a time inside that part of the day. Say it in your request
("every weekday morning") or pick it in the schedule editor.
Sources under answers: when Sentient used the web, your email, a file or a connected app to answer, small
chips under the reply show what it used: the page, the email's subject, the file or the app, next to the memories
it had in mind. Click a chip to see the part Sentient read and open the page, email or file. Answers that used
nothing show no chips.
Incognito chats: start one from the button next to New chat, the command palette or Ctrl+Shift+N
(Cmd+Shift+N on a Mac). Nothing you say in it is remembered: Sentient saves no memories from it, learns no skills
from it, leaves it out of what it learns about you and of chat search, and deletes it when you leave it or close
Sentient (or after 30 minutes without a message). The chat has a tinted header and an Incognito label so you always
know. It can still use what Sentient already knows about you; turn that off in Settings > General > Conversations.
One local model job at a time, your chats first: a local Ollama model now works on one thing at a time, so
background work no longer pushes it off the graphics card and slows your chat to a crawl. Your chat always goes
next; tasks, suggestions, memory upkeep and chat titles wait while you are chatting (and for 30 seconds after), and
while your laptop runs on battery. A small "Model busy" sign in the title bar says what the model is working on and
what waits. Change the wait, allow background work on battery or turn the queue off in Settings > Models. Cloud
models never wait.
Read only or Read and write for each connection: when you connect Gmail, Calendar, GitHub or any other app,
and on every custom MCP server, pick Read only or Read and write. Read only lets Sentient look things up
there but never change, send, delete or run anything: those abilities are hidden from the model and refused if
it tries anyway, in chats, tasks, helpers and scripts alike, even with an Allow rule or approvals turned off. Switch
any time from the app's panel in Integrations (or the server's row); it applies at once, no restart. For a server
that runs other apps' actions by name (Composio's multi-execute), Read only still allows the look-ups and refuses
the rest, and those look-ups no longer ask for approval. Fewer tools also helps small local models pick the right one.
Claude through your own Claude Code (experimental): if Claude Code is installed and signed in on your computer,
your chats can use Claude through it (Settings > Models, off until you turn it on). Sentient never reads or keeps
your Claude login, and leaves API keys and other sign-in settings in your environment out, so Claude Code always
uses the plan you signed in to. Claude Code's own tools stay off, so anything it wants done goes through Sentient's tools,
approvals and rules, and Stop everything stops it. It only answers your chats: tasks, suggestions and other
background work keep using your other models, and it can't be the memory search model. Anthropic may change how
this is counted or allowed.
Browser downloads: files Sentient downloads while using a website are saved to downloads/ in your Files
folder, and the browser result says where. In an attached browser, only downloads from Sentient's own tabs and
their popups (including target="_blank") are saved, while downloads from other user tabs remain in the browser's
native Downloads folder. Sentient's own downloads briefly pass through the browser's download folder before being moved to the downloads directory. When attached, if a download cannot be matched to a confirmed browser download record,
it returns a clear error rather than falsely reporting success.
Push to talk and dictation into any app: hold Ctrl+Alt+Shift+T (Cmd+Option+Shift+T on a Mac) anywhere, speak,
and let go: what you said goes to Sentient as a chat message, and the answer is read aloud. Press Ctrl+Alt+Shift+D,
speak, and press it again (or just pause): your words are typed where your cursor is, in any app. A small bar at the
bottom of the screen shows the microphone is on, with a level meter; Esc cancels. Speech is turned into text on this
computer. Dictation drops "um" and "uh" and fixes punctuation, or keeps exactly what you said, or lets your fast model
polish it without ever changing your words (if it does, the tidy version is used). Sentient never types into a
password box it can recognize, and puts back what you had copied. Change or turn off the shortcuts in Settings >
General, and pick the cleanup and language in Settings > Voice. Stop everything turns the microphone off.
Share your screen with a shortcut: press Ctrl+Alt+Shift+W (Command+Option+Shift+W on a Mac) to share the
window you are using, or Ctrl+Alt+Shift+R (Command+Option+Shift+R) to drag over part of the screen. Sentient opens a
new chat with the picture attached, named after the window, so you can type or say your question and send it. A
short outline flashes where the picture was taken, nothing is sent until you press send, and the picture goes to
your vision model when you set one. Both shortcuts are also in the tray menu and can be changed or turned off in
Settings > General. Since a screen can show other people's text, a chat with a shared picture checks with you
before sending anything. On a Mac, Sentient explains how to allow Screen Recording the first time.
Sign in with ChatGPT (Settings > Models > Use a plan you already have > ChatGPT plan, or ChatGPT plan in
first-run setup's cloud step): use your ChatGPT Plus or Pro plan for chat, tasks and background work. Click
Continue with ChatGPT (with OpenAI's ChatGPT logo), sign in in your browser and allow Sentient; your plan's
models then show up in every model picker, and Use ChatGPT for every job switches the main and background
models in one click (the Cloud preset uses your plan when you have no other cloud key). Main and background jobs
use the everyday model OpenAI ranks first on your plan, background jobs with the lightest thinking it allows, so
nothing is guessed from model names. After your first sign-in Sentient says once that you're using your ChatGPT
plan, and while your next message goes to a plan model the chat box shows Using ChatGPT plan with a Manage
usage link. The sign-in stays in your system keychain, renews on its own, and Sign out removes it. Manage usage opens your ChatGPT usage settings, and reaching
your plan's limit says so plainly. Memory search keeps its own model, because ChatGPT plans don't include one for it.
Models sized for your computer: Sentient checks how much memory and graphics memory this computer has before
anything is installed and recommends a local model and how much text it reads at once (for example qwen3:8b reading
8,192 tokens on an 8 GB graphics card, qwen3:14b reading 16,384 on a 16 GB one). A computer with less than 8 GB of
memory is pointed to a cloud model, with a small local model offered only for chat. Onboarding shows it as "Recommended
for this computer" with a one-click download, the "Local only" model setup uses it, and the model check-up names it
in its fixes, including a shorter context length when a longer one would push the model onto the processor.
Context meter: a small "62% of context" meter next to the message box and on running tasks shows how much of
what the model reads at once a chat or task is using. From 85% Sentient says so in plain words and suggests a new
chat or a longer context length. Nothing is blocked.
Memory sources on tasks: a task's result now shows "Used 3 memories" too: what Sentient had in mind while it
worked, from its instructions and from memory look-ups it read. The list stays with the run when it stops to ask
you something, after a restart and on Retry, and each one has This is wrong like in chat. A chat reply you
stop now shows its memories right away instead of after a reload.
"Never do this" becomes a rule. When you tell Sentient something like "never delete my emails" or "don't post
to Slack without asking me", it offers to turn it into a lasting rule: a small card in the chat ("Make this a rule?
Never: Gmail > Trash") with "Make it a rule" and "Not now". A rule keeps working in every chat, even after a long
conversation has been summarized, and Settings > Approvals & safety shows it came from your message. Nothing becomes
a rule without your click, and until you choose, Sentient checks with you before using that tool in the chat.
Said in Telegram, Discord or WhatsApp, the question comes back there too (on WhatsApp, reply 1 or 2).
Memory review: memories that come from an email, a web page, a message, a tool's result, work Sentient did on its
own, or an imported document wait for you on the new Review tab of the Memory page, with a count in the sidebar.
Each one shows where it came from and the text it was taken from; approve it, change the wording and approve, turn
it down, or approve everything from one place at once. Until you approve one, Sentient doesn't use it in chats,
tasks, suggestions or what it knows about you, and no email or web page can approve it for you. Ones you don't
review are let go after 30 days, with a note. Summaries of chats that read outside content stay out of your
other chats and out of MEMORY.md. What you say yourself in a chat that hasn't read outside content is remembered as
before.
Choose where a task's results go (Details > Send results to): the usual paired chats, this computer only, or
just the chats you pick, such as WhatsApp's "Message yourself". Results, failures, questions and plans from that
task follow the choice, and so does the Daily Brief. Jobs brought over from Hermes keep their delivery: a job that
went to WhatsApp goes to your "Message yourself" chat (once WhatsApp is linked), Telegram and Discord jobs go to the
matching paired chat, and local jobs stay on this computer.
Add values to an MCP server (Integrations > Custom MCP servers): a server that lists header or environment
names without values, such as one imported from Hermes, shows Add values. Fill them in (they go to your system
keychain, never to a settings file) and the server reconnects, so there's no need to add it again.
Report every run for script tasks: besides "Tell me when something changes" and "when the script raises an
alert", a script task can now report its output after every run. Hermes script jobs come over this way, and
Hermes monitor jobs keep reporting only changes.
Coming from Hermes? Bring your Hermes setup over in one step, from the last onboarding screen or Settings >
Advanced. Sentient looks inside your Hermes folder (~/.hermes, or one you pick) and shows what it found before
anything changes: your own skills go to Skills for review (Hermes' built-in ones you never changed are left out),
MEMORY.md and USER.md become memories and things Sentient knows about you once you approve them in Memory > Review
(with a button to remove them again),
SOUL.md replaces the personality only if you say so after seeing both side by side, scheduled jobs become paused
tasks that keep their schedule (resume one and Sentient plans it for you to approve), and MCP servers are added
turned off. Keys, tokens, sign-ins, auth.json and .env are never read or copied, so remote servers need a fresh
sign-in. Jobs Sentient can't run yet (monthly schedules, shell scripts, missing scripts) are listed with the reason.
Model presets: switch every model between local and cloud in one click. Click the model name at the top of the
window to pick "Local only" (everything on this computer), "Cloud" (your Anthropic, OpenAI or OpenRouter key for
every job) or "Mixed" (cloud for chat and planning, this computer for background work and memory), go back to a
recent chat model, or undo the last switch. Save your own setup as a preset under Settings > Models. If a preset
needs something you don't have yet, the menu offers it right there: download the model with progress, or add the key.
In Telegram, Discord and WhatsApp, /model shows the current setup and switches with a tap or a number. Per-chat
and per-task model choices still win.
Use the AI plans you already pay for (Settings > Models and the first-run setup): step-by-step help to use the
monthly API credits included with Claude Max and Team plans (paste the key, test it, then switch every job to Claude
in one click with the Cloud preset, which you can undo), Connect OpenRouter with a browser sign-in instead of copying a key, and
Nous Portal as a provider with an API key. OpenRouter's and Nous Portal's model lists show up in the model
pickers, with free OpenRouter models marked. Keys stay in your system keychain and removing one disconnects. Pro
plans don't include API credits, and apps aren't allowed to sign in with a Claude account, so the app says so.
Browser profiles: keep several browser profiles, each with its own sign-ins (for example one for posting on a
social account and one for shopping), in Settings > Browser: add, rename, delete, and "Open to sign in". A task can
use a profile (Details > Browser profile), a skill can name one (browser_profile: in its frontmatter), and the
assistant can pick one when it opens a page. Sentient can also attach to a browser you started yourself (Brave,
Chrome or Edge with --remote-debugging-port), only on this computer; disconnecting never closes it. Approvals,
purchase checks, password and code refusals and the allowed and blocked sites work the same in every profile.
Your existing browser sign-ins move to the "default" profile.
WhatsApp: link Sentient to your WhatsApp by scanning a QR code (Channels > Messaging apps), then talk to it in
your own "Message yourself" chat: text, voice notes, photos and files, replies as they are written, task results and
suggestions. When Sentient needs a yes or no it lists numbered options; reply with the number. /stop, /stopall,
/resume, /new and /help work there too. Your other chats are never read or answered. This uses the same linking
as WhatsApp Web through an unofficial app, so WhatsApp could limit an account that uses it; the connect screen says so.
Sentient's messages in that chat start with its name, so you can tell them from yours. In the installer, spoken
replies arrive as an audio file rather than a voice note.
Tasks can pause to ask you a question ("Which of these two flights?") and carry on with your answer. The question
shows on the task, as a notification with quick answers, and in paired Telegram or Discord chats with buttons;
replying to the question message there works too. A waiting task keeps waiting across restarts.
Proper documentation: getting started, privacy, architecture decision records, agent guides, verification report.
README screenshots and a new banner.
Lasting rules for apps and tools in Settings > Approvals & safety: Allow (go ahead without asking),
Ask (always ask first) or Never (Sentient can't use it). Purchases always ask.
Follow-ups: once a day Sentient notices emails waiting on your reply, and your own questions nobody answered,
in Gmail and IMAP email, and suggests a ready draft. Newsletters, no-reply senders and answered threads are skipped,
and nothing is sent until you approve it. Settings under Proactivity.
Sentient stops when it keeps repeating itself: the same tool with the same details giving the same result three
times. In chat it is first asked to try something else, then it tells you it stopped; a task fails with a clear
reason and the usual "Task failed" notification. Settings > Approvals & safety.
When a task reaches one of its limits (40 steps, 30 minutes of work, and on cloud models 2,000,000 tokens or about
$5 by default) it pauses and asks: "This task has used 40 steps and isn't finished yet. Keep going for another 40,
or stop here?" Keep going raises that limit for this run; Stop here ends it with a plain message. Time spent waiting
for your answer doesn't count. Helpers have their own token and spending limits too (1,000,000 tokens and about $2).
Memory sources: under a chat reply, "Used 3 memories" shows what Sentient had in mind when it answered: the
memories and things it has learned about you that were in front of it, and any it looked up. Each one has
This is wrong to fix or forget it on the spot, and a link to it on the Memory or About you page.
Stop everything: one control that stops every running reply, task, helper, script and browser action at once
and pauses scheduled tasks, triggers and suggestions until you press Resume, even after a restart. It is in the
title bar, the tray menu, the shortcut Ctrl+Alt+Shift+S, /stopall in paired Telegram and Discord chats, and on
paired phones. Messages you queued before the stop are not sent. It never asks the model, so nothing can
talk its way past it.
Terminal: Sentient can run commands on your computer, like git, builds, tests and scripts, in folders you
allow (PowerShell on Windows, your shell on macOS and Linux). It is off until you turn it on in Settings >
Terminal and add a folder. Each command shows you the exact command and folder and waits for your OK every time,
unless it is on your list of commands that never need asking (git status and friends) or you set an Allow rule.
Output shows live in the chat, with a Stop button; commands stop after 3 minutes by default, and Stop everything
ends them too.
Formatting disks, shutting down, deleting from the registry and deleting whole drives are never allowed. Your keys
and passwords are kept out of its commands, and work Sentient starts on its own can't run any.
Daily Brief: a few lines each weekday morning at 07:30 with today's meetings, emails that need you, tasks due
or waiting for you and the weather, plus headlines on topics you pick. Turn it on in onboarding or with Set up my
Daily Brief in Notifications. It is a normal task you can reschedule, pause or delete, it only reads (it never
sends or changes anything), and it shows at most 7 lines, each with a link and a "why am I seeing this". Thumbs up
or down on a line or a section shapes the next briefs (change your mind and the latest rating counts), any section
can be turned off, and the brief expires at the end of the day. It also arrives in paired Telegram and Discord chats,
and "read my brief" works in chat and by voice.
Evening Brief: an optional wrap-up at 21:00 every day, set up from the same card: tasks finished or failed today,
replies and emails sent, files made, what is still waiting for you and tomorrow's first events. Its own task, the
same 7-line cap, links, read-only rule and end-of-day expiry.
Remote MCP servers that need sign-in: hosted servers such as Notion's or Composio's can now be used. Add a
remote server and click Sign in to approve Sentient in your browser; Sentient refreshes the sign-in on its own and
Sign out forgets it. Servers that take an access token instead get headers (like Authorization: Bearer ...)
whose values are kept in your system keychain. A server that turns Sentient away now shows "Needs sign-in" instead
of a generic error.
Check my models: a check-up in Settings > Models and in onboarding tests each role's model the way Sentient
uses it: a short reply, a tool call (and a second tool step for chat and tasks), a JSON reply for background jobs,
thinking, context length, and whether Ollama runs the model on the graphics card or partly on the processor. Each
role gets a pass, warning or failure with a plain fix ("qwen3:4b can't call tools reliably; try qwen3:8b") and,
where the fix is obvious, a button that does it. Nothing changes unless you press it. sentient doctor --models
runs the same check in a terminal.
Stuck tasks tell you. A task that makes no progress for 10 minutes, keeps hitting the same error, or reaches a
step only you can do (a password, a CAPTCHA) pauses and says why: "Sentient is stuck on 'Book the table': the page
asks for your password. Open it to help or cancel." Choose Try again, Skip this step or Cancel, or tell it what to
do. A running task shows when it last did something. Settings > Tasks.
Catching up after sleep. Scheduled tasks missed while the computer was off or asleep run once when Sentient is
back if they are less than 12 hours late, and are skipped otherwise; never a pile of old runs. One notification
says what happened ("Caught up after sleep: ran 1, skipped 2"). Each task can choose to always run once or always
skip. Nothing catches up while Stop everything is on; it happens when you resume.

Changed

Database changes are now numbered files in sentient/store/migrations/, and the schema version is the highest of them, so two changes can no longer silently take the same version. The database moves to version 13, so the first start after updating takes a snapshot first, as with any database update. (#424)
Incognito chats leave nothing behind: when an incognito chat is deleted, the long results it saved
while reading pages or files go with it. Files you asked it to write and tasks it created stay.
A task that repeats the same failing step three times now pauses as stuck and asks what to do, instead of failing.

Fixed

Pausing a task stops it now. Before, pausing a task while it was working only stopped its next runs, and the
run already going carried on and could finish while the task showed Paused. Now the run stops before its next step
(a step already in progress finishes first, so nothing is cut off halfway), shows "Paused before it finished.
Resume the task to continue.", and Resume carries it on from where it stopped without redoing earlier steps. An
approval it was waiting on is withdrawn and asked again on Resume. A task whose agents work side by side, and a
one-step action that already started, can't be stopped part-way: the task page says so, and they finish (#408).
Approval cards name browser clicks the way the engine sees them. A browser tool's label reads the clicked
element, so clicking an ordinary button reads as a click and "Send feedback" reads as sending, while clicking
"Place order" still reads as a purchase.
A paused task stays paused when Sentient restarts. Before, a task you paused while it was working carried on
by itself the next time Sentient started, and one still being planned was planned. Now it waits: the task shows
"Paused before it finished. Resume the task to continue.", and nothing runs for it. An approval it was waiting on
is withdrawn the same way. When you resume the task, it carries on from where it stopped instead of starting
over, and asks for the approval again if it still needs it (#304).
Approval cards no longer show Sentient's notes to itself. When the browser (or another built-in app) asked
before clicking, the card's small print showed the instructions written for the model, cut off mid-word ("browser:
Click the element with this ref ..."). It now shows a short plain label when there is one, like "Purchase" or
"Unlocks a door", and otherwise the card's own line (#398).
A page watch always checks the page. "Track this page and tell me when a new event is added" could be set to
start only when something arrived in your calendar or smart home, so it never looked at the page. A check that reads
a page, a search or the news now runs on a timer: every hour, or as often as you asked (#389).
Changing a rule while an email waits to go out now counts. If you set an app or tool to Ask while one of
its emails or messages waits the few seconds before it is sent, Sentient asks you before it goes out instead of
sending it under the old rule. The same goes for Only if I asked, unless your message asked for that email.
Watching a page means watching all of it. When you ask Sentient to watch, track or monitor something, or to
tell you when something happens, the plan no longer reports on every check, and its check no longer looks at only
the page's heading or first line (so an event added at the bottom is noticed). If the plan does either, the planner
is told why in plain words and writes it again. Asking for "a report every hour" or "the headline" still works.
A free "Subscribe" button no longer reads as a purchase. When Sentient asks before clicking a newsletter,
follow or "notify me" button with no price nearby, the card now says it wants to sign you up, instead of "Make a
purchase" with "Yes, buy it". It still asks under your usual rules. A button with a price, a plan, a trial, an
upgrade or "/month" on it or next to it, or a plain "Subscribe" with nothing saying it is free, still reads as a
purchase and always asks.
Telling Sentient something changed updates the memory. After "Mira starts school in June", saying "Mira's
school now starts in July" replaces June with July instead of being ignored as something Sentient already knew. The
same goes for a new place ("Ravi moved from Pune to Delhi"), a new number or a new name. Things that are only about
the same person ("Mira has a piano recital in July") are kept as their own memories.
Page watches read the page. When you ask Sentient to watch a web page, the check script it writes now reads
the page's text. Small local models used to read a part of the page that doesn't exist, so the watch saw nothing
on every check and never told you about a change. Sentient now tells the model what a page, a web search and the
news give back, and sends a script that reads something they never give back to be rewritten.
Approvals from Telegram, Discord and WhatsApp answer only the request they were sent for. When a task asks
about one step and then another, a button or reply from the earlier message can no longer answer the newer request:
the chat says the message is for an earlier request, and the newer one keeps waiting for its own answer.
Memories in Hindi stay with the right person: a Hindi memory is no longer merged into, or skipped as a copy
of, a memory about someone else ("प्रिया पुणे में रहती है" next to "Riya lives in Pune", a Karan memory next to a
Kabir one). Names written in Devanagari are recognized and matched across scripts, so "माया" is still you when
your name is Maya. For people who write in Hindi or another language besides English, setup now suggests the
bge-m3 memory search model (1.2 GB), which finds those memories much better. Changing the memory search model
re-reads every memory once, also between two models of the same size.
Sentient can now read a long result to its end: when a web page or another result is too long to read at
once, Sentient reads the rest of the saved copy part by part instead of seeing only its start again. Reading a
file you downloaded, or that saved copy, now also counts as reading outside content, so Sentient asks before
sending anything afterwards in that chat.
A reply that fails no longer looks stuck: when the model can't answer (it is unreachable, out of credits or
turns the request down), the chat shows the error and goes back to ready at once, instead of showing "replying"
and the stop button until your next message.
The approval card no longer calls an email "posting publicly": a send like send_email is now named from
the tool's name and declared risk ("Sentient wants to send an email"). The card's word guessing never reads the
arguments' content, so an email body mentioning "post", "publish" or "tweet" cannot be misread: the only
exception is multi-execute tools (Composio), where the tool_slug/action_slug/slug keys name the action
being run (e.g. GMAIL_DELETE_MESSAGE) and are treated like tool names.
A long web or email result that Sentient saved to read later, and files you downloaded, are now marked as
outside content when Sentient reads them back, so instructions hidden in them are treated as plain text.
Learning about you and skill reviews no longer skip a message or fact saved at the same moment as the last one
they read.
Tasks no longer fail with "no final answer" after looking something up: small local models sometimes went
quiet after a memory look-up. Sentient now asks once more for the answer, and if there still is none, the run
completes with a short list of what it found instead of failing. Chat replies do the same instead of staying empty.
Editing, approving or importing a skill that needs a particular tool (for example file_read) no longer hides it
from Sentient until the next restart.
Watch requests no longer fail when a check script can't be written: if the AI's check script is rejected twice
(it loops, sleeps or uses memory, as small local models sometimes do), the task is saved as an ordinary recurring
task that the AI runs each time, with a short note on the task saying so, instead of ending in an error.
Rules from chat work with Claude Code for every model: when your fast model is Claude Code, "From now on,
never delete my emails" proposes a rule again. The check now uses the chat's own model or your main model, since
it is part of answering your message. Apps that chat through POST /api/chat also get the "Make this a rule?" card.
With approvals set to Never ask, Sentient no longer stops to ask before just looking something up (a web page,
a snapshot of the browser, a search) because something you said in the chat sounded like a "never" it offered to
turn into a rule. Until you answer that suggestion it still checks with you before changing, sending, deleting or
running anything it covers. It also reads your words more carefully: naming one tool ("don't call browser_open")
suggests a rule for that tool only, never the whole app, and a limit on the request at hand ("don't cc her this
time", "in part A") is no longer taken as a lasting rule. Settings now lists what Never ask still checks: purchases,
apps you set to Ask, sending anything once a chat has read an email or web page, and a change you just told it never
to make until you answer its suggested rule.
"Every 10 minutes", "every hour", "every 2 hours" and "every 30 min" now really repeat that often. Before, a local
model could save these as once a day at 09:00. Repeats are at least 5 minutes apart.
Check scripts written by the AI that can't work are sent back with the reason and rewritten once: scripts that use
packages that are not installed, loop forever, wait or sleep, use memory, or raise an alert where Sentient should
compare results. Scripts you edit yourself are not limited this way.
A second Sentient data folder (a test or demo profile set with SENTIENT_HOME) no longer shares the keychain with
your own: it gets its own place for API keys, tokens and sign-ins, so it can't use, change or remove yours, and a
seeded demo can't message anyone through your Telegram bot. Your own data folder (~/.sentient) keeps every key
and sign-in as it is, with nothing to redo. SENTIENT_KEYCHAIN_NAMESPACE=default shares yours on purpose. The demo
seed scripts and scripts/smoke.mjs now pause the seeded tasks and turn off messaging deliveries unless given
--keep-live.
Chats using Claude through your own Claude Code now show the context meter and the "getting long" warning at
85%: the context length is read from the matching Anthropic model, which LiteLLM knows, instead of nothing.
Background work on a laptop that stays on battery: Sentient's own upkeep (memory, what it learns about you,
dreams, suggestions, skill review, chat titles) used to wait for the charger, so a laptop used all day on battery
never ran it. It now runs on battery too, gently: while the battery is above 40%, after you have not chatted for
five minutes, in short bursts at most every 30 minutes, and never held more than two hours. At 15% or below it
waits. Tasks you scheduled always run on time, whatever the battery, and so does Dream now unless the battery is
low. The title bar says "Waiting: on battery" (or
"Waiting: battery low") while upkeep waits, and every limit is in Settings > Models > Sharing your local model.
Messages and memories saved at the same moment now always keep their order, and the daily upkeep of MEMORY.md and
USER.md no longer misses a memory saved while it was reading. Summarizing a long chat no longer skips messages
saved at the same moment as the last summary.
When a cloud model says no, chat and tasks now say why in a plain sentence and what to do, instead of showing the
provider's raw reply. For example: "OpenRouter's free models have reached today's limit. Add credits on
openrouter.ai, wait until tomorrow, or pick another model in Settings > Models." This covers rate limits, an
account out of credits, a key that was mistyped or revoked, a free model OpenRouter keeps for coding tools, a model
that doesn't exist or isn't downloaded, and a provider that is down or too slow. The raw reply stays in the log.
One long tool result no longer fills a local model's whole context. A result may now take about a quarter of what
the model reads at once (Settings > General > Conversations), so on an 8,192-token model a 13,000-character Composio
search no longer triggers "This chat is getting long" after a single step. The model is told plainly when a result
was cut, and the full result is still saved. Composio's tool search keeps its useful part instead of its first
characters: the recommended plan, the tool names, which apps are connected and the main tool's settings. A connected
tool server that sends its answer twice (as text and as data) is now read once.
Turning on "Use my Claude Code for chats" now shows the Claude Code version straight away, instead of sometimes
showing nothing until you came back to the page. Settings also says that its replies take a few seconds to start
and arrive in larger pieces than with your other models.
Claude Code no longer leaves empty scratch folders behind on Windows. Each reply's folder is removed once Claude
Code has fully exited, and folders left by earlier runs are cleared when Sentient starts.
A task that can't be planned because Claude Code is your main model now says so and what to do ("Pick a
planner model in Settings > Models"), instead of saying the AI model is unavailable.
A memory waiting for your review now shows the page or message it came from, not the error from a later action
you declined or that failed.
Adding a server (MCP) that you already signed in to from another Sentient setup on the same computer, such as a
fresh install or a test setup, no longer signs you out of it. The sign-in now remembers the server's address, so it
is kept and reused for the same address and only cleared when the address changes.
Older MCP OAuth sign-ins saved without a server address are preserved in the keychain but are no longer
automatically reused, as their server URL ownership cannot be verified. Affected users may need to sign in again
for each MCP server URL.
Local models can now use large MCP servers such as Composio, Notion or GitHub. Naming the server in your message
("using Composio...") always offers its tools, and a server with more tools than fit offers its most relevant ones,
keeping its "search" and "run" tools together, instead of none at all.
A message with a link now offers local models the page reader (web_fetch) next to the web browser, and a tool
you name exactly ("use web_fetch") is always offered (#252).
Signed-in MCP servers (Composio, for example) no longer ask you to sign in again about an hour after Sentient
restarts. The sign-in now remembers where the server renews it, so renewing works after a restart; sign-ins saved
before this fix look that up once on their own.
Local Ollama models now read 8192 tokens by default instead of Ollama's 4096, so long tasks, big inboxes
and many tools are no longer cut off without warning. Change it with models.context_length, per role with
models.context_length_per_role; it never goes above the model's own maximum, and sentient doctor shows it.
A script that Sentient runs now always gets a clear answer when one of its tool requests is refused, instead of a
dropped connection when the computer is busy.
What Sentient learns about you reads your messages in the order you sent them, even when two were saved at the
same instant.
The model check-up no longer warns about the graphics card on an 8 GB card running the setup Sentient recommends
for it (qwen3:8b reading 8,192 tokens at a time). About 90% of the model fits there, which is expected; the only way
to fit all of it is a context below 8,192 tokens, which cuts off long tasks. A bigger spill, a longer context or a
different model still gets the warning.
Saving something new about you no longer overwrites a different memory: "doesn't want files written" is kept next
to "doesn't want files deleted" instead of replacing it, and so is "doesn't want emails deleted". A memory is only
changed when the new one is about the same person and the same thing. A vaguer memory that an existing one already
covers ("your sister lives in a city", "you have a sister") is no longer added, and neither are gaps like "your
sister's job is unknown".
Everyday phrases like "If you don't know, say so" no longer make Sentient suggest a Never rule for unrelated
tools. A suggested rule for a single tool now needs your words to name an action, the app or the tool, so
"Never write files for me" still suggests blocking file writing.
Edge's downloads hub no longer replaces the active page or appears as a browser tab when a download starts.

Security

Text from web pages, emails, messages and other apps now reaches the AI clearly marked as outside content, with a
standing rule to treat it as information and never follow instructions hidden in it. A page can't fake the end of
the marked part. Your own messages are not marked. To compare, turn it off in Settings > General > Conversations
("Mark outside content").
Once Sentient has read an email, a web page, a message or anything else other people wrote, it asks before
sending, posting, inviting people to an event, typing into a web page, running code or opening an address that
could carry your data to a new site, even for apps set to Allow and with approvals off. The
approval says why ("Sentient read content from Gmail in this chat, ..."). A chat stays this way until you start a
new one. A task that read outside content pauses and asks "OK to ...?" instead, then carries on after your yes or
stops after a no.
Work Sentient does on its own (proactive checks, the heartbeat, follow-up scans, dreaming and any subagent they
start) can only look things up and change Sentient's own things. Sending, deleting, buying, running code, changing
anything outside Sentient or creating tasks is refused in code, even with an Allow rule. Sentient may offer the
refused action to you as a suggestion to approve. Chats, your tasks and suggestions you approve work as before.
One-time codes, verification codes, magic sign-in links and password reset links in Gmail and IMAP email are
replaced with a short placeholder before the AI reads them, in tool results, proactive suggestions and follow-ups.
Booking, order, ticket and reference codes, dates, prices and ordinary links are left alone, and the email itself
is unchanged in your mail app.
Purchases now ask even when approvals are switched off, and code Sentient writes can never make one. A tool set
to Never while a request waits for your yes, or while a script runs, is blocked right away.

Installers are not code-signed yet, so Windows SmartScreen and macOS Gatekeeper will warn the first time you open them.

Navigation Menu

Repository navigation

Sentient v3.0.0-beta.1

Choose a tag to compare

Sorry, something went wrong.