Changelog

Keep the whole build in view. Recent releases, newest first.

Installers for the current release are on the download page. The upcoming entry lists features still waiting on Vector's own setup; it is separate from the published version history and download links. Service setup, provider registrations and platform publication checks remain explicit gates. Installed copies of Vector receive new releases through their in-app updater.

1. Features waiting on Vector's own setup

   Upcoming — unversioned

   Features still waiting on separate setup include consent-based public sessions on Vector's service, Vector Teams, the Vector GitHub App, the public SDK and container images, and provider sign-ins needing registrations, such as xAI, Poe, DigitalOcean, GitLab and GitHub Copilot. No release date is announced for the services still awaiting setup. Version 2 is reserved for the final release.

2. Deep research preview, Vector Local, and new model choices

   October 2026 1.99.111

   Deep research arrives in preview with editable plans, Focused and Deep modes, isolated research browsers, and reports linked to source quotes. It makes no code changes. Research searches in each agent's own browser, with a fallback when search is blocked, and does not solve CAPTCHAs, disguise the browser or rotate addresses. Vector Local adds desktop controls for the model servers you choose, web access, and an origin-based request log. It is a policy inside Vector, not a firewall; programs the agent starts have separate network behavior, and Windows has no shell sandbox. Settings → Memory → About you makes learned preferences editable and undoable, and more Vector agents can use relevant preferences. Host locally keeps installed coding agents under the owner's control, with up to four running at once, explicit Try a fix actions, and exclusions from background work, learning, training and benchmark exports. The model picker separates free models provided by Vector from your own provider choices, with your own OpenRouter free copies under OpenRouter and labelled (free). Eligible new free models join the daily refresh, with training-policy consent and US only tags where applicable. The Codium catalog and pricing page now feature DeepSeek V4.1 Flash 1017, Claude Haiku 5.5, GPT-6 Luna, Kimi K3.1 Pro, and Gemini 4 Argon; DeepSeek is listed first and is the default, and Kimi and Gemini use credits faster. Model download notices clear when the model is ready. The required CLI remains 1.99.99; desktop connections to older engines explain which newer features are unavailable, and Vector Local prevents an older WSL engine from starting. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

3. The self-evolving AI agentic harness: Vector Brain learns about you, your company and each repository

   October 2026 1.99.110

   Vector now keeps a brain on this computer, one for you and one for each repository: the rules you write, your MEMORY.md, what you ask it to remember, what it learns while it works, and what your company's specs, PDFs, runbooks and API descriptions say once you give them to it. Every request but a greeting carries your rules, about the first 800 tokens of MEMORY.md, where 1.99.109 sent the whole file, and the entries you pinned, then recalls only the few lines that bear on it, within a budget set by the request's size. A greeting or a bare "ok" carries none of it unless work is waiting on it. Vector saves a short lesson, decision or preference when you state one, when a fix makes a failing check pass, or when a lasting decision is made, and uses it from the next request. Each one shows as a Learned line in the session and under Recently learned in the new Settings → Memory page, with Undo; after Undo, Vector does not save the same words, or a close rewording, for 180 days. What it learns in a session that read web pages or other outside content, ran MCP tools, or had a guest or a subagent in it waits for your review. A card or memory entry an answer cited in a run whose checks went from failing to passing gains weight in recall, and one cited in a run you reverted loses it. Add documents in Settings → Memory → Documents, or choose Remember or Remember instead on a file you attach: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), Markdown, HTML, OpenAPI, JSON, YAML, CSV and text. Vector reads each one on this computer into sections, passages and typed records, such as requirements, constraints, decisions, contracts and procedures, and brings the few that bear on a request in as short cards. A cited card shows as a chip naming the kind of record, the document and the page, and counts as verified only when Vector gave that card to the model in the session. Pictures of text, such as scanned pages, are named but not read. In a repository you trust, Vector also reads its own READMEs, docs folder, decision records and API specs. Each record is linked to the code that defines, uses or serves what it names, so a requirement comes in when you open its code, or a file that imports it, even when the request shares no word with it, and editing a file linked to a requirement asks the agent to check the change still meets it. When linked code and a document stop agreeing, Needs a look shows both sides, for example "The document says 50 MB; the code says 64 MB", with the line in the code, and it lists documents, rules and memory that disagree with each other; only you settle them. An optional local search model, a one-time 24 MB download you ask for in Settings → Memory → Search model, also finds the passages that say what a request means in other words. It runs on this computer, and until you download it, documents are searched by keywords and code links; its full-precision fallback is not yet verified, and the Code-aware model is a preview. Everything the brain keeps is stored and searched on this computer, and beyond your rules, the start of MEMORY.md and the entries you pinned, each request sends your model only the few lines that bear on it. Your MEMORY.md and personal memory (not your rules), private documents and what Vector learns from them never go to a free model labelled May train on your prompts or to a session a guest is in. Vector removes what looks like a secret from what it reads before saving it, and keeps no copy of a file you add unless you turn on Keep copies of added files; a kept copy is the file as you added it, secrets included. Let Vector learn as it works turns learning off, Read this repository's docs stops Vector reading a repository's own docs and forgets what it read of them, Tidy and erase erases what Vector learned and leaves the files you wrote alone, and a session that read outside content asks every time before a tool changes RULES.md, BRAIN.md or MEMORY.md. Deleting your Vector account does not reach the brain, because it is on your computer. Your MEMORY.md is the same file, now edited on the Memory page instead of Settings → General, and what Vector records about checks that fail in a repository, and the repairs that worked, moves from the repository's .vector folder to Vector's data folder on this computer, so a commit no longer takes it along; Vector copies an existing .vector/failure-memory.json and does not delete it, so it stays in the repository until you delete it. Free models inside of Vector are now labelled Provided by Vector, and their limit notices, which already gave a reset time, now say when they are expected to be available again. When an update cannot be installed, Vector points you to vectordev.ai/download with your free Vector account. The agent's search no longer fails when a matching line is not UTF-8 text. Vectorscope reviews in the desktop app do not get linked requirements yet; vector review in the terminal gets them, and vector memory comes to the terminal, with the next CLI release. Windows WSL servers keep running the 1.99.99 engine, without the brain, until a later desktop release requires the new CLI. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

4. Simple messages stay simple, free models on your Vector account, billing in the app and one spend limit

   October 2026 1.99.109

   A greeting or a thank-you now gets a direct answer in one small request, with no tools or extended thinking, and everything else runs with the normal tools at the model's usual effort. Subagents and high effort come in only when you ask for thoroughness, for example "be thorough" or "use subagents", or when the agent finds the work is large; the LLM-as-a-judge, when you switch it on in Settings, checks large tasks that changed files, or every task that changed files if you choose. Free models inside of Vector now cover every free OpenRouter model that can use tools, plus OpenRouter's Free Models Router, provided by Vector to anyone signed in to Vector, with no OpenRouter account of your own. Free models whose providers may train on prompts are labelled May train on your prompts and are used only after you accept once, and invited guests never reach them. Settings, Account shows who is signed in, your Vector Codium plan, its renewal and the share of this month's credits left; buying credits, subscribing, changing or cancelling a plan, updating your card and invoices open Stripe's own pages in a Vector window, and deleting your account is confirmed in Vector after you sign in again in your browser. Codium sales are still closed, so those buttons say Coming soon. Clicking your name opens Account, and when an update is ready the Update button takes the Settings gear's place, on Home too, and asks before restarting while agents are running. Spend limits are now one overall set per run, per day and per month across all your repositories: the per-repository monthly limit from 1.99.108 is dropped on update, and a repository's own budget.run is ignored with a one-time notice. Settings no longer has the rules, Appearance, Editor, Chat, Voice or Personalization pages: Vector always uses its dark look and your computer's language, Vector Tab turns on and off from the editor's status bar, and repository rules live in .vector/RULES.md, where deleting or editing a line in Vector's section removes or changes that rule. Vector calls your folders repositories everywhere. A repository with its own Vector settings asks once whether to trust it and remembers the answer, repositories created in Vector are trusted automatically and clones always ask, and until you trust one your language servers, formatters and MCP servers stay off in it. Vector Cloud now opens from Home; its button, the database banner and Publish are gone from inside a task. Vectorscope's Apply fix works and refuses to write over lines that changed since the review, an optional triageModel in .vector/review.json runs a cheaper first pass that the review model double-checks, only confirmed findings are posted, a repeated finding names every place in one comment, and a GitHub review that cannot start says which secret or model is missing. The read-only agents (explore, review, judge and security) now refuse shell commands Vector cannot fully check, such as hidden command substitutions, and other agents ask before running them; allow rules, including Always allow answers, no longer approve them. Redirects through a link that points outside the repository ask first, and a linked rules file is never read into a prompt. GitHub reviews and the terminal get these Vectorscope, sizing and shell changes with the next CLI release, and Windows WSL servers keep running the 1.99.99 engine until a later desktop release requires the new CLI. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

5. Background subagents you can steer, Vercel and Supabase controls, project trust and per-run spend caps

   October 2026 1.99.108

   Vector now starts batches of up to eight subagents in the background by default, leaving the conversation available while they work. It can explicitly wait when it needs their results, inspect their latest activity, give them new instructions at a safe step, and stop selected tasks or the whole batch. Results that finish together arrive in one notification. The Background tasks panel brings subagents, team members, swarm workers and outside agents into one view, with activity, dependencies, elapsed time and recorded spend. Talk to this subagent sends a correction to a supported running agent; Ask Vector about this adds a progress question to your conversation. Unsupported controls stay unavailable, and missing outside-agent token or cost data is shown as unknown. Individual swarm workers can be stopped, including before they start; their dependent work stops too, and completed changes are retained. Subagents can receive compact context and dependency results instead of rereading whole conversations. Scoped messages have size and rate limits, explicit file edits respect assigned paths, and child and batch spend caps stay under the launching run's Run Guard. After an engine restart, interrupted tasks keep their records and partial results and require an explicit continuation. The new vector tasks commands and the TUI's /tasks panel inspect, message and stop work; headless vector run waits for background results with a finite timeout. These terminal features need the new standalone CLI release and are not added to an already installed older CLI or WSL engine by a desktop update. Offline scripted comparisons are published with the benchmark method; they do not establish live price, speed or quality parity. Vector Cloud now runs your Vercel and Supabase accounts from the workspace, and supports only those two: Netlify, the read-only AWS inventory and Vector's own hosted publish target are removed, so Vercel is the only place the agent publishes. A Netlify connection saved by an earlier version is deleted from this computer on first launch; the access you granted Vector stays on your Netlify account until you remove it there. For Vercel: projects, deployments with Promote, Roll back and Redeploy, environment variables and domains. Publishing now makes a preview unless you choose production, where 1.99.106 went straight to production, and Vector no longer adds a Deployed with Vector badge to your site's index.html. Domains are read from the Vercel project itself; the domain records earlier versions kept on this computer are no longer read. For the linked Supabase project: the database, with the exact SQL shown before it runs, migrations, storage limits, functions, logs and security checks, and a new organization when your account has none. Vector's migration history moves from public.vector_migrations to a private vector.schema_migrations table: on the next migration run Vector copies the old rows into your Supabase database's new table and drops the public one, counts migrations the Supabase CLI recorded as applied, and stops before applying anything if an applied migration file was edited. The agent has the same controls as the Cloud screens, and Logs, Database logs and Security can hand what they show to the agent. Every change to a cloud account asks each time, with no Always allow, even in automatic mode, and that includes vercel and supabase commands in the shell and release or settings buttons on their dashboards in Vector's browser. Disconnecting a sign-in also asks Vercel or Supabase to remove Vector's access, and says so when they did not confirm it. In a project you have not trusted, Vector ignores the project's own settings, commands and skills and does not start your language servers, formatters or MCP servers there. Nothing carries over from earlier versions, so after updating each project asks once: choose Trust project in the session header to load its settings again, and changing those settings asks again. Invited guests can no longer switch on tools in their own sessions, commit, switch branches or write inside .git, and they see and add to sessions only in their own workspace. Shell commands that read a .env file now ask like the read tool, a git diff that would print files outside the work tree is checked like a file read, file writes are checked where they really land, and an MCP server's sign-in opens only web pages. Vector's own helpers, such as Vectorscope reviews, Codespace edit and autocomplete, the Browser Agent planner and the swarm planner, stay inside the screen that started them and never show up as sessions, in session lists, notifications or session counts; the session that started them shows their tokens and cost, including for stopped and failed work. The Run Guard caps what one run spends: set it in Settings, Spend limits. Vector estimates each next step and stops before one that would cross the cap, within one step for each agent running, and subagents and helpers count against the same cap. Under a dollar cap, a model with no catalog price stops after 40 unpriced steps, and Vector Codium steps count as unpriced; free models and ChatGPT plan steps are not counted in dollars. An agent that gets the same result from the same tool call, or the same error, three times in a row is paused, a provider outage ends after 10 retries or 15 minutes, and the stop card says why and offers Continue. Sessions keep stable prompt prefixes so providers can reuse their caches, ChatGPT conversations can continue over a WebSocket connection, history is downloaded again less often, titles no longer hold up the first answer, an edit and its check can share one response, and Vector loads the packages for language servers, MCP, GitLab, PowerShell and telemetry only when they are used. Background subagents and agent swarms are steadier: Stop and Discard finish even when the engine stops answering, swarm workers merge into staging one at a time, a conflicting merge is refused before anything is written and one that stops part-way is undone, Resume no longer reruns a task that had already merged, and the computer stays awake while agents in parallel workspaces or their checks run. Vector Codium is built in for when sales open: a Vector Codium provider then appears in the model picker, Vector never picks it on its own, not even for an attached image, and Settings shows the share of this month's credits left and whether purchased credits are available, running low or used up, never an amount. The /codium page lists $10, $50, $100 and $200 monthly plans and top-ups, and its buttons say Coming soon until sales open. Models included with Vector, through your own free OpenRouter account or the free models provided by Vector while they are on, now follow a no-training policy: Vector offers only free models it has checked with the provider allowlisted to serve them, every request asks OpenRouter to deny data collection, and providers may still keep prompts under their own policies, as the model picker now says. Earlier versions used only zero-data-retention providers, so Vector tells you this once, the first time you send to one of these models. The branch button in the session header is a neutral chip again, and Vectorscope's pull request header and pull request numbers look as they did before 1.99.105. The bottom of the sidebar shows who is signed in, with your Google profile photo (or your initial) and your name or email, and opens Settings; signed out, it offers Sign in to Vector. A saved Light theme, or Follow system on a light system, now applies from launch rather than only after switching it in Settings, and a server's menu in Settings opens above Settings, so a server can be edited, made the default or deleted there. Sign in with ChatGPT remains available. The npm CLI stays at 1.99.99, and vectordev.ai/install currently installs a CLI build numbered 1.99.107 from 11 October, which predates these changes, so vector trust, vector codium, vector tasks, the TUI's /trust and /tasks, and vector run --max-cost and --max-steps arrive with the next CLI release. Windows WSL servers also keep running the 1.99.99 engine this desktop requires until a later desktop release requires the new CLI, so on a WSL server the shell prompts for cloud commands and .env reads, the git diff checks, project trust, the Run Guard, loop pauses and the new subagent controls do not apply yet. The guest limits apply to vector invite from the next CLI release, because vector invite runs from the CLI. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

6. Agents keep working while your computer is idle, and Sign in with ChatGPT is back

   October 2026 1.99.106

   While any agent or subagent is running, Vector now keeps your computer from going to sleep, so long runs no longer freeze when you step away; the display can still turn off, and closing a laptop lid still puts it to sleep. When you come back to Vector after it was in the background, or after its connection to the agent engine drops, the open session and its subagent cards reload, so finished subagents no longer stay marked as running. Pressing Stop on a subagent that already finished now settles its card and says so, instead of appearing to do nothing. Sign in with ChatGPT is back: connect OpenAI with a ChatGPT Plus or Pro account to use GPT models. Vector now sends a daily usage count: how many sessions and subagent sessions ran, and, when you are signed in, token totals, recorded cost, the models and effort levels used, and streaks, with the app version and operating system. It never sends prompts, code, file names, model output, keys or IP addresses. Turn it off in Settings, General, Share usage counts.

7. Subagents keep working in the background

   October 2026 1.99.105

   Subagents now run in the background by default: the agent hands a task off, keeps working with you, and gets an automated report when the task finishes, and tasks that finish together report in one message. A message sent to a running subagent reaches it at its next step, and one that depends on other tasks waits for them. Stopping the agent's turn leaves the subagents it handed off running; stop them all from the background tasks panel. A settings change, such as a new provider key, no longer stops running subagents: Vector applies it once their work ends. Vectorscope can now review your own changes before you commit: Review my changes reviews the staged, unstaged and new files in the open project, as `vector review --uncommitted` does. Set up automatic reviews opens a pull request that adds the GitHub Actions workflow `vector github install` writes, on your agent's model, and lists the repository secrets to add; no API key is sent to GitHub, and Vector asks GitHub for permission to change workflow files only for this. A repository whose workflow reviews only when someone comments /vector review is shown as such. Branch names and pull request references now show in violet. A new install can start free: Start free with OpenRouter, in the composer and in the getting started checklist, connects your own OpenRouter account and selects its best free model, with your draft kept ready to send. The terminal agent stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

8. Vectorscope double-checks its reviews

   October 2026 1.99.104

   Pull Requests is now called Vectorscope, with the same icon, and its screen is redesigned. Reviews are more careful: after the review, a second pass re-reads the code behind every finding above a nit and drops the ones it cannot confirm, and the findings it keeps are marked Double-checked. The reviewers also read the failing CI checks for the commit under review, with the log of the step that failed, the places in your project that use the functions and types the pull request changes, and your project's rules in AGENTS.md (or CLAUDE.md) and .vector/RULES.md. A review runs on the model you last chose for the agent, and its progress line names that model; a model set in .vector/review.json still takes precedence. Comment, Request changes and Approve now post each finding on its own line, with GitHub's one-click suggested change when a fix is safe to commit; if GitHub refuses the line comments, the review goes out as one summary that lists every finding. Dismiss a finding with a reason to leave it out of what is posted, and choose Don't flag this again to save a project rule to .vector/RULES.md that later reviews follow. Fix with agent opens an isolated agent on the pull request's code with the finding as its task: it checks the problem is real, fixes it, runs the tests and commits without pushing. Reviews and merges are pinned to the commit you were looking at, so a push that arrives in the meantime is never reviewed or merged by mistake. The panel no longer reloads itself while agents are running, and its header is no longer hidden under the window's title bar. In Settings, text boxes and buttons no longer sit directly on a section's dividing line, and the search field shows one frame instead of two. Built-in Sign in with ChatGPT is turned off until OpenAI authorizes it: connect OpenAI with an API key, or use the separately installed Codex runtime. If you signed in with ChatGPT before, Vector tells you once and stops using that sign-in. The Terms, Privacy Policy and Software License now name Krishna B / VectorDev.ai and keep your rights under the LGPL for the open-source libraries Vector includes. The terminal agent stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

9. Pull Requests with your GitHub sign-in

   October 2026 1.99.103

   Pull Requests now works with your GitHub sign-in in Vector, and nothing has to be installed. Open Pull Requests, choose Connect GitHub, enter the code it shows on github.com, and your repository's pull requests appear: list, open, read the diff, create, comment, approve, request changes and merge, and run a Vectorscope review, all from Vector. Failed GitHub Actions runs and their logs come through the same sign-in, cut down to the step that failed. This is the same GitHub sign-in Open from GitHub uses, and the token stays in your computer's keychain; Sign out of GitHub in the panel switches accounts. If the GitHub CLI is already signed in on your computer, Pull Requests uses that login without asking again. When an organization requires single sign-on or approval for Vector, the panel says where to authorize it. Links you click in Vector's browser now open, including Google search results, results that pass through a redirect and links that open a new tab; before, a click that led to another website did nothing. The agent still needs your approval for each new website it visits, and when it is stopped it now names the link so it can ask you for it. Release notes and Changelog in the app open the changelog in the documentation. The terminal agent stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

10. A redesigned Changes panel

    October 2026 1.99.102

    The Changes panel is redesigned. Each changed file is now one row with its name over its folder and its added and removed line counts, under a summary of every file's totals, and a row expands into a short preview of its first change. Opening a file fills the panel with a focused reader instead of a side-by-side diff next to a file tree: one column of line numbers, a + or − beside each changed line, word-level highlights, your line comments inline, and a header with Back, the file's counts, a pager across the changed files, Open in editor, and a menu for the full file, split view (when the panel is wide enough), the previous or next change and Copy path. Back or Esc returns to the list, and a second Esc closes the panel; < and > move between files, [ and ] between changes, and Cmd+F (Ctrl+F on Windows and Linux) finds within the diff. Diffs in the conversation and in file tabs look as before. Background tasks now opens as a floating window by default, beside the Changes panel when there is room; its dock button still docks it, and Esc closes it either way. The terminal agent stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

11. Open a GitHub repository from Vector

    October 2026 1.99.101

    Besides opening a folder on your computer, you can now open a repository from GitHub: choose Open from GitHub on the Home screen or in the project menu, or press Shift+Cmd+O (Shift+Ctrl+O on Windows and Linux). Pick one of your repositories or paste a GitHub link, choose where to put it, and Vector clones it with a progress bar and opens it like Open Project. Public repositories work without signing in; connect GitHub to see your own and your organizations' repositories and to clone private ones. Vector uses your GitHub sign-in only for the clone and never writes it into the repository, its settings or Vector's logs, and a canceled or failed clone leaves nothing behind and never replaces an existing folder. Push to GitHub now lists only repositories you can push to, and on Windows Vector also finds Git for Windows installed after Vector started. The terminal agent stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

12. Messages work again in the desktop app

    October 2026 1.99.100

    This release fixes the desktop app, where in 1.99.99 every message failed with "Bun is not defined" on macOS, Windows and Linux. Local plugins and skills loaded from a URL failed in the desktop app with the same error, and plugins installed from npm and providers whose SDK Vector downloads on first use could not load there; all of them work now. The 1.99.99 apps for Intel Macs, Windows on Arm and Linux on Arm did not open, because they were built without the terminal component for their processor. 1.99.100 includes it on every platform, and a terminal component that cannot load now affects only terminals. A copy that does not open never reaches its update check, so on those computers download 1.99.100 from vectordev.ai/download with your free Vector account and install it over the old one; your repositories and settings stay where they are. On Windows, Vector's Git commands now follow your Git line-ending and symlink settings: a file whose only change is its line endings no longer shows as edited or blocks switching an agent's branch, and Vector's commits keep the line endings your repository expects. Also on Windows, Claude Code and Codex installed with npm receive the whole Parallel Workspace prompt instead of only its first line, Cursor Agent installed as a command script says it cannot take the prompt instead of running with part of it, detecting these agents no longer fails when one of them or VS Code is installed as a command script, Publish finds Vercel and Netlify CLIs installed with npm, and merge and pull-request checks run npm, pnpm and yarn instead of skipping them. A failed agent check no longer repeats in a loop while the Parallel Workspace composer is open. On Linux, Restart in 1.99.99 and in-app updates to it could open a new, empty profile; 1.99.100 opens your original profile again, so anything you signed in to or set up only in that empty profile needs to be set up again. The terminal agent was not affected and stays at 1.99.99. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

13. Vector is free, updates itself, and agents can change branch

    October 2026 1.99.99

    Vector is now free: create a free Vector account to download the desktop app, and the same account signs in the terminal agent. Installed copies now update from inside the app: choose Check for Updates and Vector downloads the new release, installs it and restarts. You can change an agent's branch: click the branch name in the session header to switch that checkout to another local branch, or type a new name to create one from the current commit. Vector never stashes or discards your work to switch, and it refuses while a Vector agent is running in that checkout; in agent workspaces Vector manages, you can create a branch but not switch to an existing one. 1.99.99 is the first release with no dependency on the upstream service. Connect a provider with your own API key, sign in with ChatGPT for OpenAI models, or use Connect OpenRouter with your own OpenRouter account; its eligible zero-price models appear under Free models inside of Vector and stop at their limits instead of switching to a paid model, and the free models provided by Vector stay off. The model list is current again, including GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol and GPT-6 Luna for ChatGPT sign-in, and Vector refreshes it from vectordev.ai when it starts and every hour. On first launch after upgrading, Vector imports your 1.99.8 settings, agents, commands, skills, plugins, MCP servers and permission rules, and keeps the originals. Security fixes keep stored API keys, auth headers and MCP tokens out of what Vector's server sends to connected clients and invite guests (an invite guest still has full access to your workspace, including the terminal, so invite only people you trust), keep Vector's internal secrets out of the integrated terminal, and protect passwordless local servers from rebinding web pages. Cloud connections keep client secrets on Vector's server, project variables keep their names when you change projects, local .env writes preserve literal values, and agent publishing uses one linked destination or asks you to choose; registration of the Cloud sign-ins is still pending. Reliability: stopping a prompt during first startup no longer leaves later prompts stuck, search recovers after an interrupted first lookup, an interrupted workspace no longer hangs shutdown, shell commands keep their final output when a process exits quickly, local plugins resolve JavaScript-style imports to their TypeScript source, and switching agents keeps the selected free model. Code Archaeology has been removed; Rewind code + chat still takes you back to earlier work. Dialogs have a look of their own, the send button lines up with the composer, the code reviewer is called Vectorscope, crash reports open as editable drafts, and providers without artwork use a neutral Vector icon. The Terms and software licence now include a minimum liability cap. This release is not yet code-signed, so a new install may show an operating-system warning on first launch.

14. Meet Vectorscope, the code review bot

    September 2026 1.99.9

    Vector's code reviewer is now Vectorscope: /vectorscope on a pull request, vector vectorscope in a terminal, and the Pull Requests panel in the app, all running the same review that reads the repository around a change. The older /vector and /vx mentions still work. The documentation is now a page per subject with search, and the landing page carries twenty feature cards. You can also delete your Vector account from the account page: it revokes your CLI tokens and asks you to type your email to confirm.

15. Subagents that work in parallel

    September 2026 1.99.8

    For big tasks the main agent launches general-purpose Subagents on its own, one per independent part and in parallel, and does small tasks itself; eight Subagent specialists keep their focus, and Settings → Agents can turn Subagents off. They show as cards in the chat and in the Background tasks panel. Turned on in vector github install, code review covers each pull request when it opens and on every push: one summary, and comments on the exact lines, each with a severity and, where it is safe, a fix you can commit from GitHub. It follows your .vector/review.md rules, re-reviews only what changed and says what got fixed, runs locally as vector review and from the desktop app's Pull Requests panel, and every review states its model and cost. The editor follows every agent live, the main agent included; Claude Code, Codex and Cursor chats answer like any other; the 16-agent limit is gone; and every model your providers offer is listed.

16. A database it can create for you

    September 2026 1.99.7

    The agent creates a real Supabase project on your own account, asks which organization when there are several, and writes the keys and client into your repository. Cloud actions that are merely unfinished now say what to connect instead of failing.

17. Cloud that closes the loop

    September 2026 1.99.6

    The agent reads your deployment logs to diagnose a failing deploy, and applies the .sql migrations in your repository to the linked database, tracking which already ran so re-running is safe. Publishing still ends with a real browser check of the deployed URL.

18. Multiplayer, Follow mode, and GitHub that ships PRs

    September 2026 1.99.5

    vector invite shares a live workspace with teammates over one link; the editor now follows the agent as it types, line by line; /vector on a GitHub issue opens a pull request with a Changes table, check output, measured cost, and the judge's verdict; and MCP servers that stop answering are recovered instead of wedging the app.

19. One workspace, two ways to work

    August 2026

    Agent and Editor now share one session, files open in persistent full-screen tabs, Settings is a searchable full-screen destination, and the Agent Dashboard, browser, terminal and review use one calmer purple project shell.

20. Scheduled work that runs while you are away

    August 2026

    Vector keeps a tray presence and stays resident after the last window closes, so a recurring task still fires with no window open. The tray shows what is armed and when it next runs, can run or pause everything from there, and notifies you when a run finishes.

21. Economics that actually learn

    August 2026

    Every session now feeds measured token and cost evidence into model recommendations, so the ranking is built from what providers actually reported rather than from list prices.

22. Plugins that install themselves

    August 2026

    Vector fetches whatever runtime a plugin needs, so Computer Use and other tools connect without a terminal detour.

23. Measured spend and in-app help

    August 2026

    Model economics now records real provider-reported tokens and cost per run, plus a Help AI Assistant and one-click bug reporting inside the workspace.

24. Agents can actually reach each other

    August 2026

    Fixed the teammate message tool, which was registered where the engine never looked, so collaborating agents could not reach each other.

25. Agents that work together

    August 2026

    Shared workspaces where agents message each other, an Agent Dashboard with clash detection, per-agent edit attribution in the editor, and AI pull request review.

26. Licensing and disclosure

    August 2026

    HEIC conversion moved to the operating system, removing the last copyleft dependency, with expanded privacy and terms coverage.

27. Vector desktop release and download refresh

    July 2026

    Verified installers for macOS, Windows, and Linux, served from the new download pipeline. Signed builds continue through automatic updates; unsigned previews stay manual-download only.

28. Review tools and shell polish

    July 2026

    A refined review panel, window recovery fixes, and a cleaner desktop chrome across the workspace.

29. Models and settings controls

    July 2026

    Restored provider management, model pickers, and execution settings throughout the rebuilt shell.