The workspace
A real code editor over the repository already on your disk, with file search across the whole project, language diagnostics, a terminal, an editor agent, and Canvas to hold them all at once.
Vector
An engineering workspace that An engineering workspace that remembers how you work, records what a run really cost, verifies its work before calling it done, drives a real browser, ships from your own cloud accounts, and runs a team of agents on the days you want one.
Vector opens a repository already on your computer and puts an editor, file search, diagnostics and a terminal beside an agent that can plan, edit, run commands, and drive a real browser. What it learns about you stays in a file on your machine. What a run cost is the number your provider reported. And on the days one agent is not enough, you can run up to sixteen.
BuildEditor and agent
VerifyChecks and browser
ShipYour own accounts
Local-first. Free models included, or bring your own. Available for macOS, Windows, and Linux.
Nine parts, one workspace
No single feature is the reason to use it.
Vector is a workspace, not a wrapper around one trick. The editor, the memory, the measured spend, the verification, the connectors, the browser, the cloud accounts, the parallel agents and the schedule all sit in the same place, over the repository already on your disk.
Local memory
What Vector learns about how you work lives in a Markdown file in your own config directory, is read in every repository, is never uploaded, and is deleted when you erase it.
Model economics
The tokens and cost a run spent, taken from what the provider itself reported. Nothing estimated. A run nothing reported is marked unmeasured rather than shown as free.
Controlled browser
A real browser view the agent drives against the running product — and one that refuses to type into password, one-time-code and card fields, handing the step to you instead.
Verified completion
Turn on LLM-as-a-judge and a separate subagent compares the finished work against your original request and the evidence, and sends it back for repair on a failed verdict.
Cloud Services
Deployments, domains with real DNS checks, environment variables, databases and logs, run through the Vercel, Netlify, Supabase and AWS accounts that already belong to you.
MCP and plugins
GitHub, Linear, Sentry, Figma, Playwright, Postgres, Stripe, Slack and more, plus any MCP server you add — and Vector installs the runtime a plugin needs instead of sending you to a terminal.
Parallel agents
When one agent is not enough, run several in their own checkouts or together in one, watch them from a single dashboard, and merge only what you trust.
Automations
Recurring work across repositories on a repeat you set, each run a real session with a transcript waiting, still firing after the last window is closed.
Codespace and Canvas
An editor, a terminal, and an agent in the same window.
Vector opens a repository that is already on your computer into a real code editor — file tree, file search across the whole project, language diagnostics from the language server, a terminal, and an editor agent in a panel beside the file you are reading. Edits an agent makes while you watch are attributed in the editor to the agent that made them, so work landing in an open file is never anonymous. Canvas arranges the editor, the browser, the terminal, notes and the cloud controls onto one surface when you want them together. None of this is a hosted IDE: the files are the files on your disk, and your repository is never proxied through Vector.
Local memory
It remembers how you work. Only on your computer.
Vector keeps durable facts about how you work — your stack, your conventions, the corrections you would otherwise repeat — in one Markdown file inside your own config directory: MEMORY.md under ~/.config/vector, or wherever XDG_CONFIG_HOME points. It is read across every project and every repository, so something you say once in one codebase is still true in the next one. It is never uploaded. Settings shows the exact path, the file's size, the date it last changed, and the full text of what is stored. Erasing it deletes the file rather than blanking it, so nothing recoverable is left behind for someone who cleared memory precisely because they did not want it kept.
Model economics
What the run cost, as the provider reported it.
Vector records what a run cost by reading the numbers the provider itself reported on the assistant messages — input, output, reasoning, cache reads and cache writes — along with the provider's own cost figure. Nothing is estimated from character counts. A run whose provider reported no usage is recorded as unmeasured rather than as zero, because a fabricated zero reads as free. Finished work builds a ledger of outcomes per kind of task, and once a model has at least three recorded runs in a category Vector can suggest it in the composer, ranked by check pass rate first, measured cost second, median latency third. Below three it suggests nothing at all: the cold start is silent rather than confident.
- What is counted
- Input, output, reasoning, cache read and cache write — as the provider reported them, plus its own cost figure.
- What is never counted
- Nothing is inferred from character counts, and a run with no reported usage is stored as unmeasured, not as zero.
- When a recommendation appears
- After three recorded runs of that kind of task. An unmeasured model never sorts ahead of one that reported a real cost.
Controlled browser
The browser is part of the loop, and it will not touch your password.
The browser is a real Chromium view embedded in the Vector window, not a screenshot service. The agent opens the running product, clicks and types, reads console and network failures, captures what it saw, repairs the code, and loads the page again. It will not type into a password field, a one-time-code field, or a card number field — that is refused where the keystrokes are executed, not merely discouraged in a prompt — and the run hands you the live browser to finish that step yourself before carrying on from what it can see. You can take the browser back at any other moment too, and everything it observed stays attached to the session that asked for it.
Verified completion
Completion you can check, instead of a confident summary.
Turn on LLM-as-a-judge and finishing stops being the agent's own opinion. Before it starts, it writes down the observable success criteria your request implies. After implementing, it has to exercise the work rather than describe it — run the tests, typechecks, build, browser or CLI checks — and capture what they actually printed. It then hands the original request, those criteria, the changed files and that evidence to a separate judge subagent, which returns PASS, FAIL, or INCONCLUSIVE. A FAIL has to name each blocking finding with its evidence and the specific repair; the agent applies them and is judged again, up to three rounds, then reports what is still blocking. It is off until you turn it on, it costs extra model calls, and it does not apply to the plan or quick lanes.
- Criteria before code
- The observable success criteria are written down before the work starts, not reverse-engineered from whatever was built.
- Evidence, then a verdict
- Tests, typechecks, builds and browser or CLI checks are run and their real output is handed to a separate judge subagent.
- FAIL sends it back
- A failed verdict names the blocking finding and the specific repair. Three rounds maximum, then the blocker is reported plainly.
Cloud Services
Ship it on the accounts that are already yours.
Deployments, domains, environment variables, databases, logs and build settings sit beside the project they belong to and run on accounts that already belong to you. Authorize Vercel, Netlify or Supabase once and Vector drives the deploy on your account, hands back the live URL, and can promote a release, roll one back, re-run its checks, or pull its runtime logs. An AWS profile already configured on this computer is read through the AWS CLI you installed. A custom domain is checked with a real DNS lookup, so “verified” means a CNAME actually resolves to the target rather than that a form was submitted. Your code is never proxied through Vector.
MCP and plugins
112 connectors, and the runtime installs itself.
Vector ships a catalog of 112 connectors you can turn on today — GitHub, Linear, Sentry, Figma, Playwright, Supabase, Postgres, Stripe, Notion, Slack, Chrome DevTools, Context7, Exa and more — grouped into categories like observability, project tracking, databases, design and search, alongside any MCP server you configure yourself. A further seventeen are listed in the catalog and marked coming soon rather than presented as ready. Several of them are published to PyPI and run through uvx, which most machines do not have. Rather than sending you to a terminal, Vector looks for that runner on your PATH and in the places uv actually installs itself, installs it through Astral's own installer when it is missing and tells you it is doing so, then rewrites the plugin's command to the resolved absolute path so it still spawns from a process that has no login shell.
- 112 ready to connect
- GitHub, Linear, Sentry, Figma, Playwright, Supabase, Postgres, Stripe, Notion, Slack, Chrome DevTools, Context7, Exa and more, by category. Seventeen further entries are marked coming soon.
- Your own MCP servers too
- Anything you configure yourself runs beside the catalog entries, with the same tool surface for the agent.
- No terminal detour
- A plugin that needs uvx gets it: resolved on PATH, installed through Astral's installer when absent, then rewritten to an absolute path.
Parallel agents
And when one agent is not enough, run sixteen.
A single task can hold up to sixteen active agents. Each gets its own checkout: a Git worktree on its own branch cut from HEAD when the project is a Git repository with a clean working tree, and an isolated copy of the project otherwise — Vector records which one it used rather than failing quietly. Or put them in one shared checkout, where they see each other's edits as they happen and message a named teammate or the whole team; you choose who may reach whom, and a message to someone you have not linked is refused with a reason rather than dropped. The dashboard shows status, model, files touched, cost so far, elapsed time, and both agents whenever two are holding the same file.
Automations
Work that repeats, whether or not you are there.
Put recurring work on a repeat — every day, weekdays only, one weekday a week, or once at a set time — and each run opens a real Vector session in that project's directory, sends your prompt, and leaves the finished transcript waiting. One task can name several repositories, each with a new session or a thread you already have, and each reports its own result so one failure does not condemn the rest. A recurring task re-arms itself the moment a run settles. This lives in the Vector desktop app: close the last window and it stays resident in the menu bar or tray while tasks are armed, and on a desktop that offers no tray it closes as it always did rather than leaving a process you cannot quit.
One Vector
One place to write, direct, verify, and ship.
Type the code yourself or hand it over, and the repository, the tools, the evidence, and the release destination stay connected either way.
Editor, file search, terminal
A real editor over your own files, with file search across the project, language diagnostics, a terminal, and an editor agent in the panel beside them.
Memory and measured spend
Local memory carries how you work across every repository and stays on your machine. Model economics records the tokens and cost the provider reported, and marks the rest unmeasured.
Review before main
Inspect diffs, focused checks, GitHub pull-request findings, risk signals, checkpoints, and an optional independent judge verdict before work reaches the primary repository files.
MCP, providers, and Cloud
Bring your own model provider, connect approved tools through MCP and the plugin catalog, and publish through the infrastructure accounts that already belong to you.
Free and BYOK — bring the agents, models, and services you already use
- Claude Code
- Codex
- Cursor Agent
- Anthropic
- OpenAI
- MCP
- GitHub
- GitLab
- Linear
- Sentry
- Figma
- Stripe
- Vercel
- Netlify
- Supabase
- AWS
- Playwright
One operating loop
From a brief to verified, shipped work.
Vector keeps the request, repository context, your own edits, agent activity, files, terminal, browser, measured usage, review, and release in one place. Enable verified completion when you want a rubric-based judge to require evidence before the task can be called done; it uses additional model calls and tokens.
- 01 Open
Attach the repository already on your computer.
- 02 Work
Edit files yourself in the Codespace, or hand a task to an agent — one, or several in their own checkouts.
- 03 Observe
Follow plans, edits, commands, browser actions, and the tokens and cost the provider reported.
- 04 Verify
Run checks, inspect the product in the browser, review the diff, and optionally require an independent judge verdict.
- 05 Ship
Merge the hunks you trust and publish through your own connected accounts.
The conditions, stated
Where a capability has a catch, here it is.
Some of what is on this page only holds under a condition. Rather than leaving those in the footnotes of nine different sections, every one of them is written out here.
Worktrees need clean Git
An isolated agent gets a Git worktree on its own branch only when the project is a Git repository and the working tree is clean. Without one — no repository, uncommitted changes, or a worktree that cannot be created — Vector provisions an isolated copy instead and records which it used in the workspace log.
Recommendations need evidence
Model economics suggests nothing until a model has at least three recorded runs for that kind of task. There is no cold-start guess, and a model whose spend was never reported never sorts ahead of one that reported a real cost.
Unmeasured is not zero
Cost and tokens come from the provider's own report for that run. When a provider reports nothing, the run is stored and shown as unmeasured. Vector will not print a zero that reads as free.
The judge is opt-in
LLM-as-a-judge is off until you switch it on, and it spends extra model calls on every prompt it covers. It cannot apply to the plan lane, which produces nothing to verify, or to the quick lane, which has no ability to spawn the judge at all.
External runtimes are yours
Claude Code, Codex, and Cursor Agent run as workspace engines, but only where Vector detects that CLI installed on this computer, and the run is signed in by the subscription already in it. Vector never asks for a second key — and teammate messaging is carried by Vector's own agent, so an external runtime shares the checkout but not the message channel.
Automations are a desktop feature
Scheduling lives in the Vector desktop app, so tasks fire while that app is running — including with every window closed, but not while the machine is off or Vector has been quit from the tray. A scheduled run uses Vector's own agent rather than an external CLI runtime.
The shell sandbox is off by default
Commands an agent runs can be confined to their workspace by the operating system's own mechanism — a seatbelt profile on macOS, bubblewrap on Linux where it is installed — by setting OPENCODE_SHELL_SANDBOX. It is opt-in, off by default, and Windows ships no equivalent, so Vector says so and runs unconfined there rather than pretending.
Cloud actions need your authorization
Vercel, Netlify, and Supabase work once you authorize them, and AWS reads a profile you already configured for the AWS CLI. Vector acts inside those accounts; it does not host your project itself and does not proxy your code.
Dictation downloads once, then stays local
Speech is transcribed by a Whisper model that runs on your machine. The model file is fetched and cached the first time, after which nothing is sent to a cloud speech service.
The sandbox flag. The operating system confinement above is enabled with the
OPENCODE_SHELL_SANDBOX environment variable. It is opt-in and off by default, uses a seatbelt
profile on macOS and bwrap on Linux where it is installed, and has no Windows equivalent — so on
Windows Vector reports that it is running unconfined instead of implying protection it cannot provide.
Only in Vector
Things that are rarely true in one place.
The difficult part is not generating text. It is keeping what you know, knowing what it cost, proving the work holds, and leaving you in control of what lands.
Memory that is one file you can delete
Local memory is a single MEMORY.md in your own config directory, read in every repository. Settings shows its path, size, last change, and full contents, and erasing removes the file rather than emptying it.
Spend that is measured, not modelled
Tokens and cost come from what the provider reported for that run — input, output, reasoning, cache read and cache write. A run nothing reported is marked unmeasured, and an unmeasured model never outranks one that reported a real cost.
A judge that can fail the work
Optional LLM-as-a-judge hands the original request, the success criteria, the changed files and the captured test output to a separate subagent. A FAIL names the specific repair and the agent goes back to work, up to three rounds.
A browser that refuses credentials
The browser agent will not type into a password, one-time-code, or card field. It pauses and hands the live browser to you, then resumes with what it can see.
Plugin runtimes that install themselves
A catalog plugin that needs uvx no longer becomes a terminal errand. Vector resolves the runner, installs it through Astral's own installer when it is missing, and rewrites the command to an absolute path so it spawns without a login shell.
Publishing on accounts that stay yours
Deploy through your own Vercel, Netlify, or Supabase authorization, manage domains, environment variables, and databases, and verify a domain with a real CNAME lookup — from a workspace that never proxies your repository.
See who edited what
The editor attributes each live edit to the agent that made it. When two agents hold the same file, Vector names the file and both agents before either merges.
Start with no key at all
Curated free models are included, so there is a capable model available before you have connected a provider or paid anyone.
Dictation that stays on device
Speech is transcribed by a Whisper model running on your machine. Nothing is sent to a cloud speech service.
Merge the hunks you trust
Review an agent's diff and merge selected hunks or files rather than accepting the whole change, with secret scanning and a restorable checkpoint before anything lands on main.
Run the other agents too
Claude Code, Codex, and Cursor Agent run as engines inside Vector, on the CLIs and subscriptions you already have — alongside Vector's own agent, the browser, the editor, scheduling, and cloud delivery in the same workspace.
Agents wired the way you want
A team can be everyone-to-everyone, a coordinator broadcasting down to workers, or any pairing you set yourself. A message to a teammate you have not linked is refused and told so, rather than dropped where the sender never finds out.
Bring your stack
One workspace for the tools you already trust.
Use Vector's provider system, connect MCP servers and source control, call supported local agent runtimes when installed, then validate and publish through your own services.
Models
Included free models, then Anthropic, OpenAI, Google, and other BYOK providers
Agent runtimes
Local Claude Code, Codex, and Cursor tools when Vector detects them installed
Protocols and source
112 catalog connectors, your own MCP servers, GitHub, and GitLab
Cloud and validation
Vercel, Netlify, Supabase, AWS through its CLI, and Playwright
A wider product boundary
Fewer disconnected surfaces between request and release.
Vector combines capabilities commonly split across an agent, editor, browser, Git client, and deployment dashboard. Capabilities change quickly, so verify current vendor documentation. A dash means Vector makes no claim either way.
View the detailed product comparison
| Capability | Vector | Claude Code | Codex | Cursor | App builders |
|---|---|---|---|---|---|
| Editor, file search, diagnostics, terminal | Built in | Terminal-first | Terminal-first | IDE-native | Managed editor |
| Memory file the app shows and can erase | One MEMORY.md, with path, size and one-click erase | — | — | — | — |
| Provider-reported spend recorded per run | Ledger, plus model ranking | Session cost readout | — | Usage dashboard | — |
| Independent judge before “done” | Opt-in judge subagent | — | — | — | — |
| Connector catalog | 112 connectors, plus your own MCP | Via MCP | Via MCP | Via MCP | — |
| Browser the agent drives | Built in, refuses credentials | Via tools | Browser + CDP | Via tools | Preview-centric |
| Deploys on accounts you own | Vercel, Netlify, Supabase, AWS | Via tools | Via tools | Via MCP | Native hosting |
| Parallel isolated agents | Up to 16 | Subagents | Worktrees | Background agents | — |
| Recurring work with no window open | Desktop tray | — | — | — | — |
Subagents
The specialists it delegates to.
Vector's agent does not do everything itself. When a task calls for a specialist, it hands the work to one of nine subagents, each with its own focus and its own permission set. Review and Security can read but never write. Judge cannot edit files at all, and only runs when you turn verified completion on.
Explore
Finds code fast
Finds files by pattern, searches code for keywords, and answers questions about how the codebase works.
General
Multi-step work
Researches open-ended questions and carries multi-step tasks through to a result.
Judge · read-only
Verifies completion
Scores finished work against the original request and returns PASS, FAIL or INCONCLUSIVE. Runs only when verified completion is on.
Debug
Finds root causes
Reproduces a failure, isolates the cause, applies a focused repair and re-checks that it holds.
Migration
Upgrades safely
Plans and stages framework, dependency, API, schema and configuration upgrades across compatibility boundaries.
Performance
Makes it faster
Measures a baseline, then optimises slow runtime paths, builds, bundles, queries and rendering against it.
Review · read-only
Reads the diff
Inspects a change for correctness, regressions, missing tests and maintainability, with file and line. Never edits.
Security · read-only
Checks the boundaries
Reviews trust boundaries, authentication, secrets, injection risk and unsafe data flow. Never edits.
Test
Writes coverage
Designs focused coverage, writes or repairs tests, runs the suite and reports failures with evidence.
Changelog
- August 2026 Scheduled work that runs while you are away
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.
- August 2026 Economics that actually learn
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.
- August 2026 Plugins that install themselves
Vector fetches whatever runtime a plugin needs, so Computer Use and other tools connect without a terminal detour.
- August 2026 Measured spend and in-app help
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.
- August 2026 Agents can actually reach each other
Fixed the teammate message tool, which was registered where the engine never looked, so collaborating agents could not reach each other.
- August 2026 Agents that work together
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.
- August 2026 Licensing and disclosure
HEIC conversion moved to the operating system, removing the last copyleft dependency, with expanded privacy and terms coverage.
- July 2026 Vector desktop release and download refresh
Signed installers for macOS, Windows, and Linux, served from the new download pipeline with automatic updates.
- July 2026 Review tools and shell polish
A refined review panel, window recovery fixes, and a cleaner desktop chrome across the workspace.
- July 2026 Models and settings controls
Restored provider management, model pickers, and execution settings throughout the rebuilt shell.
Keep the whole build in view.
Open the repository. Write, delegate, verify, ship — without leaving it.
Get Vector