v1.2.13Desktop, web and public website. Platform delivery is verified separately.
New
New English and Russian guides explain phone access to coding agents, the Codex interface and getting started for free, with a comparison overview and clear next steps.
Fixed
An unknown app address opens a helpful card inside the workspace instead of a bare “Not Found” message.
Offline computers show that work is waiting. Project dependency installation failures explain what to check; technical details remain available separately.
v1.2.12Desktop, web and account. Platform delivery is verified separately.
Improved
Downloads appear first in your account, with installation and first-task guidance. Connecting a project opens an editable request without starting work.
After you press Start and agree to Cloud coding tools, the task starts without a second confirmation screen. Required access and spending checks remain in place.
Fixed
Public website dependencies use a patched version of the site framework.
A ready Cloud model on your computer remains easy to choose when other detected agents are not installed. Platform icons remain visible in the light theme.
Windows folder failures retain bounded diagnostic codes without exposing private paths. This improves diagnosis; command sandbox failures still require a confirmed fix.
v1.2.11Cloud service and Windows desktop. macOS and Linux retain 1.2.9.
Fixed
Cloud keeps received answer text and automatically continues after an interrupted or oversized reply. Consecutive interruptions no longer end the task by themselves; incomplete commands and file changes are requested again in full. Existing credit and access limits still apply.
Provider interruptions are recorded in diagnostics instead of being mislabeled as an oversized answer. A model’s explicit refusal is preserved as its own response.
v1.2.10Windows desktop. Other platforms retain 1.2.9.
Fixed
Cloud message refusals explain availability or account access while preserving your draft. Continuing a task keeps the actual reason when Cloud cannot accept the next request.
The end of an automatic Cloud work window is shown as a pause with saved work, rather than a process crash. Repeated technical folder notices no longer clutter the conversation.
When the Windows sandbox reports that it cannot open its working folder, Cloud can use protected project-file operations. Running commands stays unavailable until the sandbox works; this fallback does not bypass file access checks or other sandbox failures.
Cloud launch errors show the computer’s redacted diagnostic report under Details, including reports from older versions. An available model does not override a blocked local executor.
The sidebar header remains reachable in short windows. On narrow screens, opening the keyboard keeps the reply and Send button visible; results and changes remain available from the task header.
v1.2.8Windows, Linux and macOS desktop apps and Cloud service.
Fixed
After a background process crashes, its unfinished reply no longer hides the next response for up to two minutes. Quitting or installing an update also cancels a pending background restart.
Cloud can read specific line ranges in project text files, including code beyond the first displayed fragment. This also works in the files-only mode. Each page shows where to continue; file access protections and the 256 KB text-file limit remain in force.
Cloud server fix for 1.2.7; no desktop reinstall is needed.
Fixed
Claude Opus 5.5 can start and continue tasks when the app includes a project file inventory or saved conversation. The fix keeps existing model access and credit limits.
v1.2.7Windows, Linux and macOS desktop apps, web workspace and Cloud service. Existing desktop apps also receive the Cloud fix without reinstalling.
Fixed
Paid Cloud tasks can continue in shorter, fully funded stages when the remaining daily allowance cannot cover a long stage. Daily limits, account access and spending caps stay in force.
Improved
The calmer sidebar, task filters and ready-result action introduced in the web workspace are now available in desktop apps too. Projects and recent tasks scroll independently; results open beside the message field.
Managed coding models receive clearer instructions to report progress in your language, keep concise task memory and resume unfinished work without repeatedly scanning unchanged files.
v1.2.6Web workspace. Desktop installers remain on 1.2.4.
Improved
A quieter workspace frame and consistent conversation width keep the task in focus. Recent tasks can be filtered by work in progress, tasks needing attention and ready results.
Ready tasks expose their result beside the message field, with access to recorded changes and checks. Project navigation and recent tasks each retain their own scroll area.
v1.2.5Cloud allowance service. Desktop stays on 1.2.4; no reinstall is needed.
Fixed
Expired and cancelled turns now release unused allowance when every attempted request has a confirmed cost, even if the computer never sent its final status. Unknown costs remain reserved.
Cloud server fix for 1.2.4; existing desktop apps receive it automatically.
Fixed
A response cut short while writing a large file no longer keeps an entire turn’s allowance reserved when its actual cost is confirmed. Continuing in smaller parts releases the unused reserve normally.
v1.2.4Windows, Linux and macOS desktop apps; VECTA Account, the public community, API and Cloud.
New
Community opens directly inside Account, with questions, ideas, showcases, updates, your discussions and private reports. Devices and settings have dedicated screens.
Every public discussion has a readable page without sign-in, its own search description, canonical address, social preview and structured discussion data. Related topics and share links help people find and pass on useful answers.
Administration now has dedicated sections for overview, activity, activation, people, shared networks, backups, workspace plans and review queues. Each administrator sees the sections their account is allowed to use.
Improved
Public discussions keep drafts during navigation and support browser Back and Forward. Account loads the forum and editor when needed; device search and retry actions keep management clear.
A sidebar on desktop and compact navigation on phones keep reports and management tools easy to find. Administration uses stock HeroUI cards, buttons, chips and tabs with the library’s default light and dark themes.
Report period, time zone and chart choices carry between sections. Direct links and the browser Back and Forward buttons return to the right view. Transitions respect reduced-motion settings; all four interface languages are supported.
Fixed
Cloud errors distinguish connection loss, allowance limits, rejected requests and invalid or oversized replies. Each has localized recovery guidance and a support reference in Details.
Gemini Cloud tasks can start when the connected app supplies a project inventory or saved conversation. The server compatibility fix also works with installed desktop 1.2.3.
Sections load independently: a failed overview no longer blocks finding a person or checking backups. Refresh also updates the review queues, with errors and retry actions beside the affected data.
v1.2.3Windows and web. Linux and macOS remain on their previously published versions.
Fixed
Large project assets no longer fail preparation at 256 MiB per file or 2 GiB per project. Independent copies and review snapshots stream file data and check available disk space with room reserved for history. Original files stay separate.
The model menu checks current availability when opened. Cloud status also refreshes while the app is active and after reconnecting, so an old limit notice does not remain indefinitely.
Preparation errors explain whether disk space is missing or the connected computer needs an update. Cloud allowance notices distinguish a service limit from the personal credit balance.
v1.2.2VECTA Account in the browser. Desktop, API and Cloud remain on their previously published versions.
Fixed
Account sign-in opens normally instead of showing “VECTA Account is unavailable in this build”. Account builds now require an API address, and the browser and connection policy use the same settings.
v1.2.1Cloud, Auto, web, Windows and Linux. macOS delivery is separate.
Improved
Accounts with individual Gemini 4 Argon access can choose it from the existing Cloud and Auto model menu. Access is enabled separately for each account; this update does not make the model available to everyone.
Individual daily credit limits are reflected consistently in the account balance and task admission. Everyone without an individual limit keeps their existing allowance.
Fixed
Models with individual access remain restricted when starting, changing or continuing a task, including shared workspaces. Knowing a model name does not grant permission to use it.
v1.2.0Cloud, Auto, web, Windows and Linux. macOS delivery is separate.
New
Cloud and Auto can save short task notes: decisions, findings and next steps. Saved notes stay with that task through context optimization and continuation, helping the agent resume without rediscovering the same details. No setup is needed.
After making changes, Cloud and Auto get a bounded opportunity to check their result before the final answer. Checks use the same tools, permissions and task allowance as the work itself.
Task details show recorded commands and browser checks, including failures and unfinished checks. A successful command means that action completed; it does not certify the whole result. If later work changes the result, earlier checks are marked as outdated.
Improved
Temporary model-provider congestion can recover automatically within the current task limits. The conversation shows when VECTA is waiting and retrying, and you can still stop the task.
Memory saved, checking the result and waiting for the provider appear in the existing task activity row. Cloud file tasks show these states beside their saved result, without extra screens or settings.
Task memory stays small and separate from the full conversation. Saving and restoring notes does not require a separate model request; only saved details are retained, so memory is not a promise to remember every earlier message.
Fixed
A retry after an explicit provider-busy response keeps the existing work instead of restarting completed actions. Ambiguous request failures are not automatically replayed as new model calls.
An unfinished result review is reported honestly and does not start another paid turn just to finish checking. Files already produced remain available when a Cloud turn stops.
Repeated update-install requests can no longer start two restarts at once. Active work still prevents an unsafe restart, and a refused installation leaves the update ready to try again.
v1.1.2Cloud, Auto, web, Windows and Linux. macOS delivery is separate.
Improved
Context optimization keeps the latest result whole when it fits after older output is removed. A short recent progress note and the latest file-write outcome can also survive later reads, within the context limit. Older details may still be omitted.
Fixed
Offline computers no longer trigger repeated automatic change checks. Slow checks are shared, and reconnection refreshes computer state so change review can recover.
A refused computer connection stops retrying the same credential. Reconnecting with a valid registration keeps queued history; temporary connection failures still retry.
Cloud file tasks use their own saved results instead of requesting changes from a local computer.
v1.1.1Cloud, Auto, web, Windows and Linux. macOS delivery is separate.
Improved
Long Cloud and Auto tasks automatically optimize context before it fills up. Older tool output is shortened toward 15% of the available context budget, while your instructions, answers and approved plans stay intact. Protected content can require more space.
Follow context optimization in the conversation while the task continues. The saved conversation stays available, and optimization does not require an extra paid model call.
Compatible Cloud workers support more context for Auto tasks. Larger project files are copied in bounded chunks with clear size limits.
Fixed
Parallel activity keeps task history ordered. Unreadable control characters in agent output no longer break history saving, and repeated connection failures retain their retry delay.
Connect OpenRouter, Vercel AI Gateway, Ollama local or cloud, LM Studio and compatible APIs with setup guidance and model discovery. Local connections stay on the selected computer.
Use attached images as references: choose style, layout or an element, mark a region and add notes. Public example links can be added to the same editable message.
Prepare a project problem report from preview, with an optional screenshot and expected behavior. Review it in the existing chat before sending.
Improved
The reference editor keeps image choices, region selection and notes together. New model connections require VECTA 1.1.0 on the selected computer.
Edit and disconnect saved model connections on desktop and phone. Compatible vision models can receive validated image attachments through the existing project-file workflow.
v1.0.2macOS, Windows, Linux, web and account.
New
Every new account receives $1 in VECTA credits at registration. This one-time welcome credit is separate from existing Founding rewards.
Improved
Queued messages have clearer text, status, attachments and controls in one scrollable panel. Long messages open by click or keyboard.
Fixed
Saved answers stay in task context instead of following later replies. On a phone, answering a question returns you to the conversation without reopening old decisions.
The file summary stays above the message box as the conversation grows. It opens the existing file review and keeps your unsent draft.
Messages wrap to the width of each conversation when several tasks are open side by side.
v1.0.1macOS, Windows, Linux, web and account.
Improved
Conversation controls, task menus and progress rows use compact spacing and consistent text sizes.
Follow public text as it arrives in supported Cursor sessions, completed messages from Amp and tool activity from Gemini. Other supported agents show activity when they produce output, without revealing private content.
Local agents and VECTA Cloud receive clearer guidance to work in small, verifiable stages and share factual updates. Update timing depends on the agent; this release does not add Cloud token streaming.
Fixed
Usage limits stay readable with several connected computers: provider summaries keep their own space, and missing quota information uses a clear status.
Sidebar project and recent-session lists scroll independently, with subtle edge fades to show more items and keep switching within reach.
Project names and descriptions retain their contrast when you hover over gallery cards.
4 moreShow less
New tasks retain the selected local model and supported reasoning level when you change projects and open the conversation. Available Codex reasoning levels are taken from the selected connection.
Open a verified update straight from Settings, retry the failed download, and see why it failed. If the system cannot open the installer, VECTA stays usable and lets you try again.
Manually downloaded updates must match both their published fingerprint and file size before they can be opened.
Unfinished public replies clear when you stop or switch a task, and completed replies appear only once. One unreadable history entry no longer hides the rest of a desktop conversation.
v1.0.0Windows, Linux, web and account. macOS delivery is separate.
New
Open a task in its own desktop window with its conversation and unsent draft; return to the workspace when you need the full controls.
Administrator activation metrics show observed cohorts, completed observation windows and confirmed onboarding steps; attribution follows analytics consent.
A redesigned compact workspace: task board and list, focused and parallel conversations, shared model picker and project appearance.
6 moreShow less
Review files side by side, request changes and commit accepted work locally.
Connect your own CLI agent or compatible model endpoint from the connections screen.
Cloud model details show reported token rates, context and output limits with their source and checked date.
Write an optional summary for a public task receipt and review the exact preview before publishing.
Undo the last agent reply in a direct folder with a checkpoint and protection for later manual edits.
Set your first password through email verification, dictate in supported browsers, and connect a custom MCP service.
Improved
Usage limits include other connected agents and official Antigravity and TRAE guidance, without invented quota percentages.
Task plans show connected steps, their recorded progress and timing; the compact board and inbox include useful task descriptions.
Agent questions, handoffs, context and usage limits follow the compact layouts on desktop and phone.
1 moreShow less
Project setup explains how to bring in a repository or connect a computer. Profile actions share a compact, grouped menu.
Fixed
Parallel tabs reuse an already refreshed session before requesting another renewal. Answer timing follows its own run and stops before later review or apply actions.
Moving a task to another desktop window clears its original draft only after the destination is ready. A computer limit no longer blocks opening an existing workspace.
Long model names and Cloud category counts stay readable in the phone model picker.
3 moreShow less
New tasks retain the selected model, reasoning level and attached reference images through dispatch and retries.
A temporary connection problem lets you retry your saved sign-in without entering your password again.
Task history stays ordered when saving needs a retry, and repeated commit requests keep the same result.
Founding now has 500 places. Cloud preview remains available for new accounts after the first 2,000 shared turns; existing credit and spending limits still apply.
VECTA Team: one pool of machines for the whole team. A task goes to a free machine that fits its project — or waits its turn in the queue, with its place shown — and the Machines screen shows how busy each one is. You can still pick a machine yourself.
Add a Linux machine with one command: copy it from the Machines screen, run it there, and the machine joins the pool by itself with its system, memory and agents detected.
Sign a machine’s agents in from the browser: the machine’s card shows which agents are signed in, and Sign in shows the agent’s own code and link right there — no SSH.
7 moreShow less
Create a project from a GitHub repository in the browser. VECTA suggests a setup script and machine requirements from the repository’s files, and you can add secrets; their values are never shown again after saving.
One conversation for the team: every message shows who wrote it, you see who is typing and who is looking at the task, and each task has one person leading it who can hand it over.
Waiting for the team: every question and approval from every agent in one list, with Allow and Deny right in the row. The first answer counts, and everyone else sees who answered and when.
When a machine goes offline during a task, the work saved after the agent’s last reply can continue on another machine with Continue on another machine.
Roles for people in a workspace — owner, member and viewer — and an audit log of who changed what.
Plans: Free, Pro, Team and Enterprise. Pro and Team are paid from VECTA credits on the account page, under Plans & billing, which shows what is charged and when. Team can be tried free for 14 days by teams of up to 5.
A new page about VECTA Team on the site, in English and Russian.
Improved
Existing shared workspaces receive a 30-day Team trial and a private Personal space for each member. Existing personal workspaces keep Free, or retain their previous terms for 60 days when the new limits would affect them. Projects and history stay.
A new home page shows how your coding agents keep working while you are away: a day with VECTA, where agents can run, a comparison with similar tools, all four plans and a try-it-yourself approval demo. English and Russian now have their own addresses, and four comparison pages are added.
v0.38.2Web, account, Linux and Windows.
Improved
You can change a running task’s approval mode in its Run settings, including turning approvals off, and it applies to that task right away. Before, the setting only affected new tasks, so a task already running kept asking.
The model picker shows each model once under its full name, has one Default instead of Auto and Default side by side, and scrolls as one list with nothing cut off.
Fixed
When an agent’s own provider account runs out of money, such as Claude Code’s Anthropic balance, the task says so plainly — that it is that account’s balance, not your VECTA credits, and how to top it up — instead of a limit that “resets”. The provider’s message no longer appears in the conversation as the agent’s reply.
“Don’t ask again” on a command request names what it covers, such as mkdir, cat. Writing a file inside the task’s folder now qualifies too, so an agent creating files asks once per task instead of once per file. Anything outside the folder still asks.
Improved
A task in your own folder starts right away, in the folder, with no copy made first — however large the folder is.
The task screen is calmer: the agent team’s settings moved into Run settings and the team appears in the conversation only when it proposes helpers or runs them; project context moved to the ⓘ in the header; the shared-files and evidence cards appear only when there is something in them.
Run settings is one tidy list: computer, file access, the task’s approval mode with a link to the default for new tasks, spending and the agent team.
2 moreShow less
Command requests are compact: the command, what it touches and three answers, without repeated lines.
A screenshot the agent takes sits in the conversation as part of its message, and opens in a viewer that fits the window, with Fix here beside Close.
VECTA Cloud tasks in larger projects no longer pause again and again without saving anything. Each turn starts knowing the project’s files, finds the lines it needs, reads only those and changes part of a file instead of rewriting all of it, so a change takes fewer steps.
v0.37.0Web, account, Linux and Windows.
New
No computer needed for a first task: without one, New task starts in VECTA Cloud — describe what you need and press Start, and a project is made for you. The computer and AI screens, and the desktop app when it finds no agent, offer the same.
A Getting started card on the Tasks board leads to your first result in three steps: where AI works, a first task, its result. It disappears once you have one.
Backup AI: choose one in project settings, and when the task’s AI reaches its usage limit, the backup continues the same task once. The task says when and why it switched.
1 moreShow less
A task whose computer is offline can be started again on another computer that has the project, with your request already filled in.
Improved
A task shows which project decisions its AI was given. The project context screen says plainly what reaches agents: approved decisions go to every new task, the first 20 and up to 400 characters each, while discussions are for people.
Fixed
Tasks in VECTA Cloud now follow the project’s approved decisions too.
After every decision is removed from a project, a task started again no longer keeps the old ones.
Commands in work steps are shown exactly as they ran, including backslashes inside quotes, and the Founding offer no longer says “1 seats left”.
v0.36.1Account, Linux and Windows.
Improved
VECTA Cloud on your computer works with the project’s files before its one-time sandbox setup. When a command is needed it says which one, and Fix it on the AI screen lets it run commands from the next message.
On Windows, the account page offers the installer, which keeps VECTA up to date by itself, beside the portable build. From this version a portable copy updates by installing VECTA once.
Fixed
Claude Code signs in the way it does in your terminal. If its settings name a relay, a proxy, Bedrock or Vertex, tasks use them instead of stopping with “Not logged in”, and a token from claude setup-token works in folder tasks on Windows.
Every step an agent takes in a task can be opened: what it ran, what it printed, how long it took and whether it failed. Steps are named after what actually happened instead of “Worked on task”.
Automations open on an overview of your routines with ready templates — an evening summary, a dependency check, failing tests, risky changes, release notes and docs drift. A template fills in the form; nothing runs until you save.
A project page now shows where the project lives, its tasks by state, the AI it uses, its decisions and instructions, routines and who can see it, with actions for each.
Improved
Task screens are calmer: the result is split into clear cards, approval requests and problems read as one card each, the agent team is one card with its helpers, and the header controls sit on one line.
Tasks side by side are separate rounded panes, and the active one is clearly marked.
Reasoning effort is a row of named levels instead of a slider. The microphone sits inside the message field. While an agent works, Send adds your message to the queue and Ctrl+Shift+Enter sends it now; an answer to an agent’s question goes straight to it.
2 moreShow less
Usage & limits opens on AI limits, and its lists share one row shape. Menus everywhere use the same row height and highlight.
The desktop app finds a new version within minutes and as soon as you return to it or wake the computer, without a restart.
Fixed
A task whose computer is offline names that computer properly and says what is kept there.
The Founding offer stays in one place in the sidebar instead of floating over the screen and the account menu.
v0.35.0Web, account, Linux, Windows and macOS.
Improved
Screens line up on one grid: record, settings, inbox and account pages share one column width and header height, and the panel beside a record lines up with its header.
Section headings use sentence case, and buttons, search fields and view switches in toolbars share one height.
Inbox titles, groups and tasks start on one edge; fact lists, side-panel rows and the computer’s agent list have even spacing.
Fixed
Working states, mint text and accent notices are readable in the light theme, and the GitHub logo is visible on light backgrounds.
Notices stay in their corner on the new-task screen instead of covering the Discard button, and the project picker shows its full label.
Change review navigation is translated, and the Russian sidebar shows “Входящие” in full.
1 moreShow less
VECTA Cloud is marked as recommended on the computer screen again.
v0.34.1Web, account, Linux, Windows and macOS.
Fixed
Security hardening: nothing an agent leaves in its copy of a project can make VECTA run it outside the agent’s sandbox when VECTA reads the work with Git.
Dangerous mode and shared-folder teams now answer only to the person who owns the computer. Others can still stop or pause the work.
A finished task can no longer use your connected accounts, and a forum sign-in stays a forum sign-in.
3 moreShow less
The desktop app moves to Electron 44.7.0, keeps its log and saved sign-ins private to your account, and explains plainly what letting a task drive a test window allows.
Support logs upload reliably with a daily allowance. Links written by an agent show where they really lead, and hidden characters in code are marked.
The website’s privacy settings keep product analytics off unless you turn them on.
Optional product analytics help us understand which features work well and where errors occur. Choose during registration or change your choice in settings. Collection starts only with your consent; session recording is off.
Improved
Task rows, screenshot previews and run settings are more compact and consistent.
Fixed
Analytics consent follows your account across the app and account pages. Turning it off stops collection immediately, including when saving the preference needs a retry.
v0.33.0Web, Linux, Windows and macOS.
New
Agent teams can work in the current project folder or in separate copies. Choose the mode before starting: in the current folder, helpers see each other’s changes immediately and you review the combined result in the parent task. Stopping the parent also stops its helpers.
Improved
A consistent mint identity and compact HeroUI interface across tasks, projects, computers, settings, sign-in and account. Quieter surfaces and clearer typography make long sessions easier to follow.
Agent teams keep their full workflow in the new interface: the parent proposes agents and models, you approve assignments, and child conversations, results and controls stay beside the parent.
Model selection opens beside the composer with search, effort controls and reported pricing. Usage, subscription limits and task budgets have compact, readable summaries.
1 moreShow less
A denser task list, bounded board columns and contextual review actions keep large workspaces manageable. The website now presents the updated product with a complete startup workspace.
Fixed
Approval cards stay fully visible, computer shortcuts no longer overlap, and the chat composer keeps project, model and run settings within reach. Nested provider menus remain open while their form scrolls.
Codex can use VECTA’s scoped team tools in a task. Helper assignments still require your approval.
Keyboard focus returns correctly from nested menus and remains usable while forms are saving. Password reset has a clear way back; expanded dialogs keep their actions reachable.
1 moreShow less
Unavailable file comparisons retain the selected file and offer a retry. Long task names, offline states and connection errors use consistent layouts.
Website
Improved
The VECTA website has a quieter mint design, a clearer product walkthrough and screenshots from a complete startup workspace. The English and Russian pages adapt to desktop and mobile screens.
v0.32.0Web, Linux and Windows. macOS builds are released separately.
New
Enable Agent team in task settings. The parent chooses from ready, compatible connections and explains each assignment. Review the prompt, agent and model; select the helpers you want to start.
Four concurrent helpers by default, up to sixteen in settings. Each works in a separate copy. Helpers cannot create more teams; repeated starts are bounded.
Child conversations can open in the right-hand panes. The parent can read their results, send clarifications and request pause or stop. Closing a pane does not stop its task, and changes are never merged automatically.
1 moreShow less
Teams support isolated Claude Code, Codex and Z.ai Coding Plan connections, plus scoped OpenRouter and OpenAI-compatible model connections on the same computer. CLI agents without verified isolation, Cloud, direct-folder tasks, Dangerous mode and shared task budgets are not supported in teams yet.
Improved
Compact rows keep the task name, status and actions aligned. Results and team details expand when needed; file comparisons no longer fill the conversation with large cards.
Dangerous mode has a shorter confirmation: acknowledge the risks and enable it. The outside-project access warning remains explicit; plan review and questions still need your answer.
v0.31.1Web, Linux and Windows. macOS builds are released separately.
Fixed
Usage & limits now distinguishes unreadable limits from an agent that cannot run. Open the affected AI directly to see its setup or repair; unavailable data no longer incorrectly asks you to update VECTA.
Older Claude Code and Codex installations now offer an update when their CLI is verified to lack required task permissions. Failed checks remain separate; project permission boundaries stay enforced.
New
Connect Windsurf through its official Devin CLI, with a bundled logo and account sign-in instructions. Tasks use CLI mode and show output in the task terminal; command approvals remain in Devin.
Improved
Cursor CLI has an official setup guide, readable terminal output and a capability check before tasks start. Both CLI connections use the provider’s own account sign-in.
v0.31.0Web, Linux and Windows. macOS builds are released separately.
New
Connect Z.ai GLM Coding Plan as a separate agent powered by Claude Code on your computer, with GLM choices, its own logo and the same VECTA approvals. Subscription key and model access are checked on the first task.
Improved
VECTA 0.31.0 brings a compact dark interface, conversation-first navigation and consistent controls from task setup to computer pairing.
A task header keeps the title, current status and useful tools in one row. Open its context to check the project, agent, computer and file access. On a phone, tools move into a labelled menu.
Menus, dialogs, disclosures, first-run guidance and computer pairing share compact spacing, neutral surfaces and readable type. Keyboard focus, touch targets and reduced-motion behavior are part of the same pass.
Fixed
The board uses quiet status badges and consistent filters. Forms keep your draft after a failed save, explain how to retry and prevent duplicate actions while a request is pending.
Keyboard focus stays inside dialogs and returns to the next control when choosing AI. Handoff actions remain visible on short screens.
v0.30.0Web, Linux and Windows · macOS build follows separately
New
Mark an area in a task screenshot, describe the change and add the image to your message draft. Visual feedback appears only where screenshots and a compatible agent are available.
Choose two saved screenshots for a before/after comparison. Matching sizes support a slider; different sizes stay side by side.
Improved
One component system across the workspace: navigation, menus, settings, project and automation cards now share consistent controls, spacing and selection states. Your familiar screen arrangement stays in place.
Chat questions, access requests, confirmation dialogs and the message composer use the same visual language as the rest of VECTA.
A ready result can offer an independent review with another agent. Preparation explains code checks and relevant tests, adding interface inspection when visual context is available. Reviews use a separate copy and need your confirmation.
1 moreShow less
The original task links to its separate review and avoids repeating the offer for the same result.
Fixed
Chat cards keep their action buttons visible, the automation schedule switch has a full-size track, and volume settings show one clear value.
New conversation records no longer replay unchanged browser ownership over a newer live control lease.
VECTA automatically limits Cloud access for corroborated automated account campaigns. Account access and local agents remain available.
Improved
Send now options are more compact, and activity icons align with their text.
Changed-file comparisons have a clearer card with separate counts for compared, unavailable and unchecked tasks.
Fixed
The desktop app restores your saved sign-in after restarting instead of incorrectly reporting that VECTA is unavailable.
If tasks cannot load, VECTA explains how to retry and keeps the last loaded list without claiming it is empty or repeating the connection warning.
An offline-computer message names the task’s computer and offers Open Computers instead of showing a private-address placeholder.
v0.29.2Chat interface · included in 0.29.3
Improved
Chat access requests and their confirmation dialog use the same style as the rest of the workspace, with clearer actions and permission details.
v0.29.1
Improved
Agent work has expandable recorded steps, a live answer cursor, clearer decision cards and consistent task rows and message fields.
Code blocks show line numbers and a language header. Copy keeps the exact code without line numbers.
Fixed
Opening VECTA in a new browser shows the sign-in form without an incorrect session error. A previously ended sign-in still explains why you need to sign in again.
VECTA Cloud keeps working while your sign-in renews. Browser tabs coordinate renewal so opening several tabs no longer interrupts your session.
When a model stays overloaded, VECTA explains what happened and how to continue. Your saved files remain available, and the busy attempts are not charged.
Improved
Queued tasks say why they are waiting. A task with no available computer stops after one hour; a task waiting for an offline computer stops after three days. The prompt is kept for another attempt.
Founding rewards from a busy shared network may wait seven days before they can be spent. Administrators can review shared-network activity in Analytics.
v0.28.1
Improved
The new-task screen shows only what stops this task, with the one button that fixes it, and opens by itself when something does. Task suggestions have a calmer look with a short note under each.
When a folder is too large for a snapshot, the task offers to continue without one: the work goes on as usual, you just can’t roll the files back automatically.
Fixed
Signing the desktop app in through your browser no longer signs that browser out. If your sign-in there is older than 15 minutes, the page asks for your password instead.
On Windows, the window uses the system’s own minimise, maximise and close buttons, and menus and dialogs no longer slide under the title bar.
Project previews start more reliably: the port VECTA gives a project’s server is no longer taken by another program in the meantime.
Open a pull request from a finished task. VECTA pushes the task’s branch with the Git sign-in already on that computer, opens the pull request on GitHub, and shows its checks beside it.
Automations run on a schedule: every day, on weekdays, on chosen days or every few hours, in your time zone. Before you save one, VECTA shows exactly what will run, on which computer and when.
VECTA tells you when a task is waiting for you: a system notification from the desktop app, and an email when a task you started has waited 10 minutes without you opening it. Choose what you get in Settings → Notifications.
Improved
Connect GitHub by pasting a token on the Skills screen. Services VECTA can’t work with yet are marked as not available instead of looking connected.
Recorded checks show the tests an agent actually ran, with their real result, and the counts when the test runner reports them.
In the browser, the microphone explains that dictation works in the desktop app and links to the download.
Fixed
When VECTA Cloud on a Windows computer needs its one-time setup, the new-task screen says so and offers Fix it instead of asking you to download VECTA again. Computers that haven’t updated keep Cloud when the server changes.
Codex and Claude Code are no longer reported as outdated when they were only slow to start: VECTA checks them one step at a time, waits longer and shows what actually happened.
A task queued on an offline computer says it is waiting for that computer, and the desktop app uses the new vecta.build addresses.
The Analytics tab is available only to accounts with explicitly granted administrator access. Follow daily registrations and tasks over 7, 30 or 90 days in your chosen timezone.
See how people move from registration to their first task, and compare task creation, starts, completions and current statuses.
Find a user by name or email and open their activity details, including the titles and statuses of their 20 latest tasks.
Improved
The web account has clearer navigation, typography and spacing on desktop and phones, in both themes and all four languages.
Dangerous mode now activates supported agents’ native CLI permission bypass, including on later turns. Confirm the updated warning before starting a new task: some native modes bypass the CLI sandbox and can act outside the project with your computer account’s access.
Existing tasks keep their original permissions. Withdrawing control stops a native Dangerous agent before another turn can begin.
v0.26.1
New
Dangerous mode automatically answers VECTA action approvals for new tasks after a warning and explicit confirmation. Set a default in Settings → Conversations or choose it for one task; project boundaries and spending limits still apply.
Fixed
Screenshot references in agent messages open the saved image inside the task instead of showing raw link markup.
VECTA Cloud receives the exact selected model identity when answering questions about itself.
Cloud turns continue with the task board tools enabled; strict tool and permission checks remain in place.
v0.26.0
New
A task board for work across your projects: plan tasks without an agent, link project tasks to a shared goal, and follow five stages from Later to Done.
Quick-install the TakeOne guide from Skills to help an agent record a website demo when you ask. Installing the guide does not start a recording or download a browser.
The optional ThreeUI guide helps you explore a 3D scene for a suitable website. It asks before adding one and offers a static alternative.
Improved
Delegate a board task, answer questions, give approvals and review the result from its card. An agent finishing its response leaves the result for you to accept.
Use a board or list on desktop and a compact list on your phone. Live updates keep your unsaved edits and focus in place.
Fixed
Registration refuses known temporary inboxes and asks for an email address you keep. Signing in to an existing account is unchanged.
v0.25.10
Fixed
VECTA Cloud on Windows works with projects in your user folder: commands run in the project and can save files there.
No more false “helper file was removed” warning on Windows start: each installer carries only the helper for its own processor.
v0.25.9
New
Claude Code, Codex and every other agent work on Windows: where a separate copy of the project is not possible, they work directly in your folder after you agree, with approvals on and a snapshot first.
VECTA Cloud runs on Windows computers too, in a Windows sandbox set up once with one administrator prompt.
Improved
The agents screen shows working agents first, says in plain words why one did not start, and lets you copy the details for support.
The desktop app connects to VECTA’s main server, so it signs in again even when an older address stops answering. You stay signed in after updating.
Installers download straight away in your browser, and dialogs open above every task when several are side by side.
v0.25.6
Improved
Every approval can now be given for the whole task: the card says exactly which similar actions it covers, and you can stop allowing them at any time.
v0.25.5
Fixed
VECTA starts on Windows again: it no longer carries helper files for other systems that a virus scanner could remove.
VECTA Cloud writes big files much faster, so large requests like a whole browser game finish instead of running out of time.
v0.25.4
New
Work on up to eight tasks side by side. Drag a task from the list to any edge, move and resize panes, pick a layout, and share the arrangement as a link. Desktop app and web.
Watch VECTA Cloud work: each file it reads or writes and its notes between steps appear as it goes.
Improved
VECTA Cloud keeps going on bigger jobs: longer turns, and when a turn runs out of time it saves the files and carries on by itself.
Keyboard shortcuts for panes, shown right in the menus and the command palette.
Fixed
A Cloud task that ran out of time no longer loses the files it had already written.
VECTA Cloud says plainly when you ask it for a screenshot it cannot take.
v0.25.3
New
A VECTA Cloud page that explains how it works, your credits, how to earn more and which models you can use today.
Copy a support code from a task’s menu to send to the VECTA team. It holds only ids and versions, never your work.
Tasks show which model did the work, with its own logo.
Improved
Settings are split into clear sections, each with its own link.
The AI list groups what is ready, what needs attention and what is not installed.
A cleaner task menu with plain names for sharing and who can see a task.
Fixed
VECTA Cloud can write whole files in one go; if a reply is still cut off, it carries on in parts and tells you what it was writing.
Screenshots in WebP and GIF reach your other devices, and a screenshot kept on an older computer says how to bring it over.
v0.25.2
New
Every account gets its own picture, the same in the app, on the account page and in the community.
Faster, cheaper VECTA Cloud models, and one you can use for free every day.
Improved
Inbox groups what needs you by kind, and each item shows its folder, agent and computer at a glance.
Sharing a task receipt is one compact step, and the public receipt page looks the part.
Fixed
Connection and workspace notices float in a corner instead of pushing the window around, and a brief restart no longer says your workspace is gone.
v0.25.1
New
Confirming your email now gives you your Founding seat and Cloud credits right away — no extra step.
Share your invite or a task receipt in one click to X, LinkedIn, Reddit, Telegram, WhatsApp and more, with the text ready.
Improved
A much simpler account page with your VECTA Cloud credits front and centre.
VECTA Cloud is listed first wherever you choose an AI.
Fixed
VECTA Auto keeps going through a brief network drop instead of getting stuck as busy.
First-run screens are centred, and the Founding offer no longer pushes the window down.
v0.25.0
New
Let an agent without its own isolation work directly in a folder you choose, after a clear warning. A snapshot is saved first by default, and restoring it never overwrites your later edits.
Answer an agent’s questions right in the chat: pick one or several options, write your own answer, or dismiss the question. Works with Claude Code, OpenCode, Codex and VECTA Cloud.
Sign in to the desktop app through your browser: approve this computer on the account page, no password typing in the app. Both sides show the same code so you can check the request is yours.
2 moreShow less
VECTA Cloud Founding Builders: the first builders to claim a seat get Cloud credits for supported models, and earn more for their first finished task, invites and shared task receipts.
Share a task receipt: a public page with the facts of one finished task, published only when you choose and removable at any time.
Improved
Credits pay only for VECTA Cloud models. Claude Code, Codex and your own API accounts keep working exactly as before.
A cleaner homepage, and VECTA on X is now linked from every page and the community.
Fixed
On Linux, VECTA Auto project commands run inside their sandbox again; if a required system tool is missing, VECTA names it and how to install it.