Key Zemu AI Agents updates are collected at a practical level for teams operating AI workers.
v1.2.54
AI Workers can operate apps on your computer, plus a simpler sidebar
2026.10
Turn on Computer Use in Zemu Desktop and AI Workers can operate apps on your Mac or Windows PC from a simple chat request. The sidebar has also been consolidated into five items.
The sidebar is now consolidated into five items: AI Workers, Knowledge, Dashboard, Templates, and Settings. Skills, triggers, and channels are available under Settings, and old URLs and bookmarks automatically redirect to the new screens.
Deleting a skill from Worker settings now shows a confirmation dialog explaining that the change affects every Worker.
Sign in with Zemu, plus more precise numbers in Knowledge databases
2026.10
External apps can now offer "Sign in with Zemu" so you can log in with your Zemu account. Knowledge databases also pick better column types and handle numbers more precisely.
When you ask AI to create a Knowledge database, it now chooses column types such as number, date, select, email, and URL based on what each column means and the data in it. Phone numbers and postal codes are kept as text so leading zeros are not lost.
When you create a number column in a Knowledge database through AI or the API, you can now set the number of decimal places (0 to 10). Long numbers such as amounts keep their full precision when you edit or paste them on screen.
AI Workers can operate your everyday Chrome on Windows too
2026.10
From "Request an action" in Zemu Desktop, you can now hand browser tasks in your everyday Chrome to AI Workers on Windows as well. Before form submissions and similar actions, you can review the details and approve them.
You can stop a running operation from the web. Stopping keeps the connection to Chrome, so you can move straight on to your next request.
The approval screens on the web and iOS are easier to read. The Approve, Reject, and Stop buttons stay easy to reach even for long content, and secrets such as passwords are not shown on the regular screen.
Meeting recording on Windows Desktop, and account deletion requests from the web
2026.10
Zemu Desktop for Windows can now record meetings. You can also request account or organization deletion from the web.
You can now request deletion of your own account from Profile settings, and deletion of an organization from Organization settings (owners only). You can check the status of a request and cancel it from the same screens. Submitting a request does not delete any data by itself, and you receive an email when a request is accepted or canceled.
Desktop sign-in now uses the same screen as the web. After you sign in with Google or email, you are taken straight back to Desktop.
Clearer guidance when meeting recording needs attention
2026.10
When meeting recording in Zemu Desktop runs into trouble, the reason and the next step now appear at the top of the screen. Zemu also tells you about the recording time limit.
A single recording can last up to 2 hours 30 minutes. After 2 hours, Zemu lets you know that 30 minutes remain, and recording stops automatically at 2 hours 30 minutes.
When you run out of credits, Zemu shows that transcription is paused and that it will resume from the next audio once you add credits. You can add credits right from the screen.
Only the member who created a meeting can start or stop recording and regenerate the AI summary, so other members can't stop a recording by mistake.
Invoicing app setup with progress, and documents linked by name
2026.10
Setting up the invoicing app now shows its progress, and quotes and invoices created by the assistant are linked by document name.
Quotes and invoices created or updated by the invoicing app's assistant are now shown as links with the document name instead of a URL. Click one to open that document.
When the assistant finishes its work, the document list automatically refreshes to show the latest state.
From Zemu Desktop (macOS), hand browser tasks to an AI Worker using the tabs and sign-in state of the Chrome you already use.
Choose an AI Worker from "Request an action" in Desktop to open a chat with the request pre-drafted. Nothing is sent automatically, so you can review and edit before sending. If the Worker's permission is off, "Open settings" takes you straight to that Worker's Permissions & approvals.
New Workers can start without pre-registering sites or devices. When needed, add site restrictions or pin a specific device from advanced settings. Every action is checked against the user, organization, and device permissions, and other people's devices or incognito tabs are never used.
If the connection drops while the AI is responding, Zemu checks the saved state and reconnects automatically so work can continue. Completed actions are never run twice.
Pick an image for each AI Worker, plus usage hints in chat
2026.10
Choose one of 40 characters or upload your own image for an AI Worker. The chat input now shows hints about handy ways to use Zemu.
When creating a Worker or from "Change image" in its settings, choose one of 40 built-in characters (32 people, 8 pets) or upload your own image (JPG / PNG / WebP / GIF, up to 5MB). Existing Workers keep their current image or icon until you change it.
The chat input shows rotating hints such as @mentions for sub-agents and Knowledge, "/" commands for skills, and "/plan" or "/goal". Hints adapt to the selected Worker or skill and are available in Japanese and English.
Restore deleted Workers, save files to Knowledge via MCP, AI meeting summaries
2026.10
Deleted AI Workers can be restored for 7 days. Files can be saved to Knowledge via MCP, and meeting minutes are summarized by AI.
Deleted AI Workers are kept for 7 days and appear when you select "Deleted" in the list filter. Users with edit permission can restore them with their settings and sharing intact. Triggers that were stopped at deletion stay stopped after restore, so re-enable them if needed.
Regular-size files can now be saved directly to Knowledge via MCP. Saved files inherit the permissions of their destination. If another upload is using the capacity, the status shows "Waiting for capacity" with guidance to retry.
Meeting minutes now get an AI summary organizing the overview, key points, decisions, and action items, with remarks marked "Important" prioritized. Transcription costs 2 credits per 10 minutes of recording, and the AI summary is charged once per meeting and not charged if processing fails. You are notified before starting if your balance is insufficient. Triggers that cannot run due to insufficient credits are now recorded as "Skipped" with the reason, and the next scheduled run is kept.
Change text color or switch to a code block from the menu that appears when you select text in Knowledge.
The menu shown when you select text offers 8 text colors with reset, and a toggle to switch to or from a code block. The active button shows with a white background so the current state is clear in both light and dark themes.
If you select everything and turn it into a code block, pressing the button again restores the original formatting.
Page titles in browser tabs and a faster accounting Worker
2026.10
Opening a Knowledge page or database row sets the browser tab title to that page. The accounting Worker that fills in counterparties for Money Forward journal entries is faster too.
Opening a Knowledge page or database row sets the browser tab title to that page or row title, so multiple tabs are easy to tell apart.
The accounting Worker that fills in counterparties for Money Forward journal entries now fetches entries in one batch, reducing processing time and credit usage.
Upload multiple files and folders to Knowledge at once
2026.10
Upload several files or nested folders to Knowledge in one go. AI Workers can now point you to screens inside Zemu by URL.
Upload multiple files or nested folders together. Each file shows its result and preview, and if some fail, the others are still saved. Multiple attachments on page bodies and database rows are supported as well.
AI Workers can now point you to screens inside Zemu by URL, such as settings, Workers, conversations, Knowledge, and skills.
Multi-condition customer filters and self-extending Worker runs
2026.09
Filter customers by industry, source, sales status, campaign, and owner in combination. AI Workers extend their own run limit when work remains.
Search customers by combining industry, source, sales status, campaign, and owner, so you can narrow down to the customers you need to check.
When adding or editing a customer, pick attributes from existing options or enter new ones. Sales status can be changed directly from the list, and the last-updated time is shown. The form shows attributes in two columns on desktop and one column on mobile.
The initial run limit for AI Workers is 50, and the Worker extends it by 50 at a time when work remains. The reason, progress, and remaining work for each extension are shown in chat. No new approval is needed to extend, while existing explicit approval settings are still honored. The limit of 50 is not a cap on total runs or usage charges.
Goal-based Task tracking and smarter document reading for Workers
2026.09
Tasks are tracked by the original request, success criteria, and verification results to avoid unfounded completion. Workers avoid re-reading material they have already read.
Tasks are tracked by the original request, success criteria, verification results, and open items rather than as a checklist of steps. Scope is preserved when you request changes midway, which prevents unfounded completion. Earlier completion records are marked as lacking verification details.
Workers still read enabled Skills and required material once, but avoid re-reading what they have already read and unnecessary exploration, moving on to execution and verification sooner. Saved generation instructions and Custom Skill settings are supported.
GPT-6 Astra / Sol, Claude Opus 5.5, GLM 5.3, Grok 4.7, DeepSeek V4.1 Flash, and Qwen3.8 Max were added to the model list.
The model list now includes GPT-6 Astra / Sol, Claude Opus 5.5, GLM 5.3, Grok 4.7, DeepSeek V4.1 Flash, and Qwen3.8 Max. Older models are hidden from new selections, and saved Worker and chat settings are kept. Availability follows your account permissions and balance.
Compare how customers respond to shared links by industry, source, deal stage, campaign, and sales owner.
Compare shared-link engagement over the last 90 days by industry, source, deal stage, campaign, and sales owner. See customer count, customers shared with, customers who accessed, access count, and access rate.
Manage category fields on the customer create, edit, and detail screens. Page through long category lists, and check what is counted and how it is calculated right on the screen.
Share customer links, review AI Worker weekly reports, retain Worker memory, and embed Knowledge databases. Accounting retries are also improved.
Create a shared URL for each customer in the customer management App, view access history, and notify the person responsible when the link is accessed.
View and deliver AI Worker weekly reports, and retain business information in Worker memory. Accounting Workers now save document-processing progress and evidence, with improved retries after interruptions and guidance when additional authorization is needed.
Embed a database in a Knowledge page and create references between pages to navigate related information more easily.
Knowledge database charts, invoicing, and purchase order OCR
2026.09
Filter, aggregate, and chart Knowledge database records. New offerings support quotes, invoicing, payment tracking, and purchase order OCR.
Filter database rows, sort by multiple conditions, and view footer aggregates. Chart views support period-based totals, targets, and finalizing actual results.
Open the AI assistant from a database record detail screen to ask questions and take actions using that record as context.
A new AI-assisted app supports creating and sending quotes and invoices, tracking payments, reminders, and recurring invoices. The purchase order OCR template reads and processes documents, including image-only PDFs.
Faster chat input and continuous Discord conversations
2026.08
Start drafting new chats sooner, continue Discord conversations in one thread, and undo queued-message deletions.
Replies in regular Discord channels continue in the same Zemu thread, preserving the conversation context.
The new-chat input appears before the AI Worker finishes loading. Send once the Worker is ready, and open a new chat directly from the home page when signed in.
Delete queued messages without a confirmation dialog, then restore an accidental deletion with one click.
Better Knowledge editing and more Free plan integrations
2026.08
Improved block and table controls make Knowledge easier to edit. Connect apps such as Gmail and Slack on the Free plan.
Move blocks and tables using drag or keyboard controls. Table row and column operations and font-size editing are improved, with settings preserved after saving. Use the plus button to search for headings, tables, and other blocks to add.
Connect Composio-supported apps such as Gmail and Slack on the Free plan, then use them with AI Workers and event triggers. You can start adding an app even with no existing connections.
Signup and organization invitation flows no longer create unnecessary accounts. The default customer success Worker is available when the account is created.
Knowledge storage visibility and template integration icons
2026.08
Check Knowledge file usage, limits, and remaining space in organization settings, and see which services official Worker templates use.
View Knowledge file usage, the 100 GB limit, and remaining capacity in organization settings. Uploads, Markdown updates, and restores from trash also check the storage limit.
Official AI Worker template lists and details show integration icons for services such as Slack, Money Forward, Discord, and Chatwork, so you can check the required connections before choosing.
Zemu Desktop local agents are now available in all environments, including production.
Review deleted Knowledge pages and database rows in a shared trash and restore them with one click.
Deleted Knowledge pages and database rows are collected in a shared trash, where you can review and restore them with one click.
When the accounting AI worker is unsure how to proceed, it now files an escalation ticket to the Knowledge DB first, preventing missed or delayed handling.
Sales AI team template and role-based CS Knowledge
2026.08
A new resident sales AI team template works safely behind approvals, and CS Knowledge for inquiry handling can now be assigned per role.
Added a resident sales AI team worker template: drafts and records are automated, while email sending and important CRM changes always require approval.
CS Knowledge for inquiry handling can now be assigned per role, so each role references only the information it needs.
Message queue, question stepper, and thread forking
2026.08
Queue your next instructions while the agent is still running, answer multiple questions in one stepper, and fork threads into new work.
Queue your next instruction while the agent is running; it is sent automatically when the run finishes, and queued messages survive closing the screen.
Review and answer multiple questions from the agent step by step, with in-progress drafts saved and restored.
Fork a thread to branch off new work, and run local agents in production from the Desktop app.
Review each Worker's outcomes — completion rate, handled time, cost, and estimated ROI — on a dashboard.
Check each Worker's weekly outcomes — completion rate, estimated coverage of handled time, actual cost, estimated ROI, and more — over any period on the dashboard.
PM AI worker setup is simpler: the progress-note destination is now determined automatically.
A refreshed model catalog and unified sub-agent management
2026.07
The model catalog has been refreshed with GPT-5.6 added, and you can now add and manage sub-agents all from the permissions screen.
Refreshed the model catalog. We added "GPT-5.6" and tidied up rarely used models so you can pick the right model for the job.
You can now add and manage sub-agents (connected agents) all at once from the permissions screen, with tighter integration with approval settings.
When a sandbox is recreated, your workspace is restored automatically, and images embedded in knowledge pages are managed as page-owned assets, so references are less likely to break.
Slash commands, live Mermaid diagram previews, text color and highlights, and more make the knowledge editor far more expressive and easier to use.
Added slash-command formatting, live preview for Mermaid diagrams, text color and highlights, image resizing, and drag-and-drop file uploads.
Type "@" in the chat input to search and select knowledge, then attach its content as context for the conversation.
Version history now shows the date and editor's name, and each thread remembers the AI model you last used and selects it automatically when you resume.
Global AI assistant and knowledge editing improvements
2026.07
A new AI assistant drawer you can open anytime from the right side of the screen. External knowledge edits are now more responsive too.
Added an AI assistant drawer you can open anytime from the right side of the screen. It automatically understands the page you're on and supports you based on your work.
Edits made to knowledge from outside the editor are now reflected instantly, making everything feel snappier.
When a permission error occurs with the Google Drive integration, clear guidance for reconnecting is now shown.
A new mode lets you share your goal before the AI starts working. You can also copy files from the file menu.
Added "goal-first mode" — share your goal before the AI starts working. Your intent is understood more accurately, so you get better results with less back-and-forth.
You can now copy files from the file menu, making it quick to create a new file based on an existing one.
Knowledge database tool results can now be reviewed in a dedicated right-hand panel. Signed-in users can also browse the top page freely.
Results from knowledge database tools now appear in a dedicated right-hand panel, where input, output, and execution status are visible at a glance.
Signed-in users opening the top page are no longer auto-redirected into the app until they click the login link, so the landing page and announcements stay browsable.
The result panel and landing-page browsing path are separated so users can keep context while working.
Knowledge databases have been redesigned into an easy-to-read table grid with nine field types and inline schema editing from column headers. Files can now be previewed before upload.
Redesigned knowledge databases into a table grid. Field types expanded to nine (adding long text, checkbox, single select, multiple select, URL, and email), and you can rename fields, change types, edit options, and delete columns directly from column headers.
You can now preview a file's contents before confirming an upload. The file-selection display is also fully localized.
Opening an empty folder now shows cards for creating a folder, a database, or a file, so even first-time users can start creating content without getting lost.
Added a Secret save request that lets AI safely guide you to the save screen instead of pasting sensitive values into chat. AI workers and the CLI can download knowledge files.
Added a "Secret save request" so that instead of pasting sensitive values like API keys into chat, an AI agent safely guides you to the Secrets settings screen — opening a form in the right panel with the Secret name pre-filled.
AI workers and the CLI can now download the contents of knowledge files. Deleting knowledge is now unified under full-edit permission; view/edit-share users cannot delete.
The permission boundary was tightened so secret handling and knowledge-file access are safer to operate.
Lightweight knowledge editing tools and auto-saved browser profiles
2026.05
AI agents can now append or partially replace knowledge pages without loading them in full, browser-profile login sessions are auto-saved, and Slack multi-account routing no longer triggers duplicate runs.
Added the `knowledge_append` and `knowledge_replace` lightweight knowledge editing tools so AI agents can append to or partially replace knowledge pages without loading them in full — making edits on large pages much faster.
Browser-profile login sessions are now auto-saved without needing to press "Save and close". Sign in once before delegating to the AI and subsequent runs start from that logged-in state.
When the same Slack workspace is linked to multiple Zemu accounts, a single Slack message no longer triggers multiple agents by mistake — Zemu now sends one disambiguation notice when several candidates match.
Continuation threads for scheduled triggers and the new /loop command
2026.05
Scheduled triggers can now run as continuations of the same project thread, and you can create or update triggers right from chat with the new `/loop` command. Duplicate runs of the same trigger are skipped automatically.
Scheduled triggers now support a "Use continuation project" switch. When enabled, each scheduled run continues as a follow-up to the same thread.
You can now create or update scheduled triggers directly from chat by typing `/loop`. The AI confirms frequency, and when a specific time is given, it also detects the timezone.
When a previous run of the same trigger is still in flight, the next run is now skipped automatically — duplicate concurrent executions no longer happen.
Role-based knowledge permissions and a public embed widget
2026.05
Knowledge access can now be granted by role, and AI agents can be embedded into external sites with a new public widget.
Knowledge access can now be granted by role — view-only, append-only, or full edit — and public share links support issuance, revocation, and login-required settings.
A new public embed widget lets you drop AI agents into your own site or service page. Configure allow-listed domains, the model used, and usage limits to publish safely.
Editing Markdown tables inside knowledge has been refined — adding, removing, and reordering rows and columns is now much more intuitive.
Browser profiles and stronger knowledge permissions
2026.05
AI workers can now keep browser login state via attached profiles, with new knowledge permission controls and the ability to hide past tasks.
Attach a browser profile to an AI worker to persist login state and cookies during browser tool runs. A new agent tool also lets workers list, create, and assign profiles themselves.
Knowledge permissions v2 adds shareable links, view-only access, and per-worker access grants so you can control exactly what each worker can read.
Past tasks can now be hidden from the sidebar, with the hidden state synced per account across tabs.
The Help Center is now organized by category, with additional UI improvements across navigation and worker setup.
The Help Center was refreshed with Japanese and English guides for models, AI workers, skills, knowledge, triggers, integrations, billing, notifications, and member management.
Settings now links directly to the Help Center, and the sidebar has a more compact, easier-to-scan layout.
AI worker creation, worker icon editing, knowledge list display, and model selection guidance were refined.