Navigate Vector Tap the menu button to close

Open a repository, direct agents, inspect their work, and ship from one desktop workspace.

Vector is available for macOS, Windows, and Linux.

Privacy

Privacy Policy

How Vector handles account information, local work, AI requests, connected services, diagnostics, and usage counts.

Effective October 17, 2026

1. Scope and contact

This Policy explains information handling for the Vector website, desktop application, command-line tools, and Vector-operated account, support, and cloud endpoints. “Vector,” “we,” and “us” refer to Krishna B / VectorDev.ai, the operator of those services. For privacy questions or requests, email krishnabharadwaj0521@gmail.com.

Vector keeps much of your work on the computer running the application. AI requests, connected tools, help, and publishing involve other systems. The data sent depends on your instructions, selected model, enabled features, permissions, and connected accounts. Independent providers also process information under their own terms and privacy policies.

2. Account and sign-in information

Supabase provides account authentication and stores your email address, account and authentication identifiers, sign-in provider, and profile information supplied at sign-in, such as your name. If you use Google sign-in, Google processes that authentication and supplies the permitted account information. We use account information to authenticate you, verify access, and respond to account requests. CLI access may use an account-linked token for verification.

Website sign-in does not itself upload your repository or local conversation history. Authentication uses browser storage to keep your session signed in and complete sign-in redirects. Signing out or clearing that storage removes the local session but does not delete the account.

3. Local repositories, conversations, and memory

Repositories, session history, tool results, preferences, checkpoints, workspace metadata, and local logs are generally stored on the computer running Vector. The embedded browser can retain cookies, sign-in sessions, and other website storage on that computer. Connected remote runtimes or servers may instead hold the data they process.

Optional local memory stores notes and preferences in a local configuration file. You can inspect and erase that file in Settings. Relevant memory and instructions can be included in requests to your selected model. Deleting local memory does not delete a copy already sent to a provider, copied into conversation history, or included in a backup.

Dictation. With microphone permission, desktop dictation processes recorded audio on your computer using a local Whisper model. The model is downloaded and cached; its hosting service receives the download request, but this transcription path does not upload your audio. Browser dictation uses the browser's speech-recognition service, which may send audio to a remote provider under that browser or provider's privacy practices. The resulting text enters the composer and, when submitted, is handled like a typed prompt. Stop recording or revoke microphone permission to prevent further capture.

4. Models and AI request content

To answer a task, Vector may send prompts, conversation history, selected or agent-read files, code, images, instructions, memory, repository metadata, and tool output to the selected model service. Tool output can include terminal text, search results, browser content, or information returned by an integration. Content is selected according to the task and enabled permissions; it may include personal or confidential information present in those sources.

Model requests use the provider you connect and its service and account, except for Vector Codium and the Free models inside of Vector that Vector provides, described below. Vector no longer uses the former third-party model gateway. Requests can pass through a configured gateway or proxy. Each recipient's retention, abuse-review, and model-training rules depend on the service and your agreement or settings with it. Vector does not promise that every provider offers zero retention or excludes all request content from training.

Free models inside of Vector and OpenRouter. When you choose one of the Free models inside of Vector, whether provided by Vector or through your own OpenRouter account, your request content goes to OpenRouter and to the model provider serving that model. Free models provided by Vector pass through Vector's server to check your account and limits, and send OpenRouter a one-way hash of your Vector account identifier with each request. Vector does not request zero data retention for these models. Most of their providers keep prompts and outputs under their own policies; OpenRouter records Phala, which serves Xing4.0, as keeping none. The models fall into two groups:

  • No-training models. By OpenRouter's records their providers do not train on prompts. Each request to them asks OpenRouter to deny data collection.
  • Models labelled “May train on your prompts”. Their providers may log your prompts, code and outputs and use them to train their models, or Vector has not yet reviewed their policy and treats them that way. From desktop 1.99.109 and the next CLI release, Vector labels these models and asks you once before sending your first prompt to one; nothing is sent to them until you accept. You can withdraw that choice with Withdraw on the Free models inside of Vector card in the desktop app's Help panel, or with vector models --decline-training in the terminal. Earlier versions of Vector are offered no-training models only. Don't use these models with private repositories, secrets or personal data.

Vector itself never trains on your prompts, as the Training paragraph below says. Each model's group is on our Models and providers page. This relies on the providers' published policies and OpenRouter's records, not an independent audit, and does not change your own OpenRouter account settings, which can hide models that may train.

Vector Codium. When you choose a Vector Codium model, your request content goes from Vector to Vector's gateway on vectordev.ai, which Vercel hosts. The gateway checks your sign-in, your Codium plan or purchased credits and the credits held for your requests still running, then forwards the request to OpenRouter on a key Vector holds for your account, not a key of your own. Each request asks OpenRouter to route it only to providers OpenRouter records as zero data retention and to deny data collection; this relies on OpenRouter's records and the providers' published policies, not an independent audit. The gateway passes request content through without storing it. With each request it sends OpenRouter a one-way hash of your Vector account identifier, not your email address or name, so that use can be attributed to one account.

For Codium, Vector keeps a billing record in its server-side data store: your Stripe customer, subscription and billing-period identifiers, your top-up purchases, and the encrypted OpenRouter key for your account with its credit limit. The credits held for a request in progress are kept for about eleven minutes at most. To buy, Vector creates a Stripe customer with your email address and your Vector account identifier; Stripe processes your payment details, which Vector never receives, and keeps the payment records. Deleting your Vector account closes open Codium checkouts, cancels active Codium subscriptions and disables your Codium keys; the billing record is kept, marked closed, as section 12 describes.

Training. Vector does not use your prompts, code, or model output to train any model, including any model Vector builds itself. What a model provider does with the requests you send it is governed by that provider, as described above.

5. External agents, plugins, and connected tools

If you select an external runtime such as Claude Code, Codex CLI, or Cursor Agent, Vector can pass your task and working directory to the installed program and display its responses. That program may read repository files, use its own credentials, transmit context, and maintain its own logs and history under its own permissions and provider policies.

Plugins, MCP servers, repository services, and other integrations receive the requests and data needed for the actions you authorize. Responses can enter the local conversation and subsequent model requests. A standing permission or scheduled task may allow multiple operations without a new prompt each time. Disconnecting an integration prevents future use through that connection but does not remove records the provider already received or revoke credentials everywhere they may be used.

Live workspace invitations. A live workspace invitation contains guest credentials. People who have the link and can reach your server can access the workspace through the guest account; a tunnel provider may also process connection traffic. Share invitations only with trusted people and stop the server to end live access.

6. Browser and computer use

Browser automation can inspect page text, URLs, page structure, screenshots, and technical browser output, and interact with sites you open. Those sites receive ordinary browser traffic and any information submitted to them; their cookies and privacy policies apply.

Where available and enabled, computer-use features or optional plugins can list available targets, capture a selected window or screen, and send mouse or keyboard input. These features process window or application names, target and frame identifiers, screenshots, and requested input actions. Operating-system permissions, including screen recording and accessibility permissions on macOS, may also be required. Availability and controls depend on your platform, application version, and plugin.

Screenshots and action results can be stored as session content and sent to your selected model provider. They may reveal notifications, messages, personal information, or confidential content visible in the selected area. Close or hide unrelated sensitive content before use. You can stop the session, disable the feature or plugin, and revoke operating-system permissions. Stopping future capture does not delete previously recorded session content or copies already received by a provider.

7. Cloud connections and published content

Cloud features can transmit repository files, build output, deployment settings, database operations, resource identifiers, and configuration to providers such as Vercel, Supabase, GitHub, or GitLab, depending on the connection and action. Responses, resource metadata, and logs may be displayed or retained locally and included in agent context when you use them in a task.

Some connections use Vector-operated OAuth endpoints to exchange authorization codes, refresh access, or revoke a connection. Those endpoints process the necessary codes, access or refresh tokens, and transaction metadata before returning the result to the application. Credentials are therefore not confined to your device throughout every connection flow.

When you publish, Vector uploads the repository's files and deployment settings from your device to the Vercel account and project you connected and linked; Vector does not host the site. A published deployment can be reached by anyone with its URL unless Vercel's deployment protection covers it, so publishing can disclose content publicly. Remove private data and secrets before publishing. Local deletion, sign-out, or account deletion does not itself remove a deployment from your Vercel account; remove it there. For content published through Vector's earlier hosted publishing service, which has been retired, contact support to have it removed.

Vector-hosted public sessions. Where this optional service is available, publishing a session stores its visible transcript in Vector's Supabase service and makes it readable by anyone with the link, without signing in. The transcript includes its title, visible prompts, assistant responses, displayed reasoning, code, and tool input and output. Review it before publishing: these fields may contain personal information, confidential material, or secrets, and Vector does not guarantee automatic removal of sensitive content. Account credentials, application settings, and hidden system instructions are excluded as separate data fields, but credentials or other sensitive information appearing in the visible text can still be published. Attachments are represented by descriptive labels; the public viewer does not fetch the attached files or their URLs.

Public sharing requires explicit consent. You choose whether future updates to that conversation are public and select an expiry no later than 30 days after publication. If you separately choose to remember account-wide consent for automatic sharing, future sessions can be published when automatic sharing is enabled, including sensitive text they may contain. Repository automatic-share settings alone do not authorize uploads. Disable automatic sharing to stop new automatic publications and use Unshare to remove existing public copies. Sharing is optional and is not required to use Vector.

Successful unsharing, link expiry, account-access revocation, or account deletion stops public access through Vector. Successful unsharing and account deletion erase the live transcript; expired live transcripts are erased by scheduled cleanup. If a removal request fails, the link may remain public; retry it or contact support. Deleted or expired content may remain in service-provider backups until those backups expire under the configured retention policy. Contact us for the applicable storage location and backup-retention information. People who have already viewed a transcript may keep their own copies, which Vector cannot erase. The viewer asks search engines not to index the page, but this does not make a public link private or prevent others from sharing it.

To prevent delayed upload requests from restoring removed links, we retain a minimal deletion record indefinitely without the transcript. It contains the random share identifier, a hash of its management secret, and lifecycle metadata, linked to your account identifier until account deletion clears that link. These records are not returned by the public viewer.

Earlier third-party shares. The former third-party public-share service is no longer used for new publication in Vector. This does not remove copies published there by earlier releases. Removing a Vector-hosted public session, deleting local history, or deleting your Vector account does not remove those historical copies. Their removal requires the former service's controls or separate cleanup, and recipients may also retain copies.

8. Help and support

The in-app help assistant sends your recent help conversation and matching Vector documentation through a Vector-operated endpoint to Groq to generate an answer. It does not automatically attach your repository or coding-session history. Anything you type or paste into the help conversation is part of the information sent. Service error logs may include upstream error details.

When you submit a bug report, we receive the description, any optional reply email, application version, operating system, architecture, release channel, and technical context included by that report flow. Resend delivers the report to our support mailbox. Email you send directly also becomes support correspondence. Avoid including credentials or unnecessary confidential information.

9. Diagnostics and service logs

Remote crash diagnostics are off by default. If you enable them and reporting is configured, Sentry receives application version, platform, stack traces, and technical error information. Vector disables default personally identifiable information and performance tracing and removes request, user, breadcrumb, and extra fields from diagnostic events before sending them. Error text and stack traces can still contain incidental information. Repository files, prompts, terminal output, and credentials are not intentionally attached. You can disable future reporting in Settings. Local crash dumps and application logs remain on your computer unless you separately share them.

Our website, API, hosting, and security providers process network information such as IP addresses, request time, browser or client details, requested routes, and response or error information to deliver and protect the service. These operational records are separate from optional desktop diagnostics. Update and download requests also disclose ordinary connection and requested-release information.

Help and bug-report abuse controls use keyed identifiers derived from network addresses and, for some reports, a derived email identifier. These identifiers are pseudonymous, not necessarily anonymous. The help counter expires after ten minutes; bug-report network counters expire after one hour and email counters after 24 hours. Expiration of those counters does not delete separate provider logs.

10. Usage counts and website analytics

What is counted. To learn how many people use Vector, whether they come back, which versions and platforms to support, and how much model use runs through Vector, the desktop application and the command-line tool report usage counts to Vector. The desktop application sends a random install identifier created on your computer, your account identifier when you are signed in, the application version, the operating system and CPU architecture, and the number of sessions and subagent sessions active that day (in UTC). It reports shortly after it opens each day and again, checked about once an hour, whenever something it reports has gone up. The command-line tool's daily account check, which already identifies your account, also reports the tool's version, operating system and CPU architecture at the first command of each UTC day, and counts that day as a day of command-line use.

Model use. When you are signed in, the desktop application's reports, and the command-line tool once each UTC day, also send the totals shown in the desktop application's Usage & streaks settings, worked out on your computer: token totals by type (input, output, reasoning, and cached) and the number of model responses; the model cost the application recorded; tokens, tasks, and recorded cost for each of the last seven days with any use; the models and providers you use most (up to five), as their identifiers with each one's tokens and share of your tokens; the reasoning effort levels used, with their tokens, responses, and share; the number of chats completed and conversations; your active days and current and longest streaks; and your average and longest task time. The command-line tool sends these with a random install identifier that Vector does not store, because its counts are kept against your account.

Usage counts never include prompts, code, file names or paths, model output, provider keys or credentials, IP addresses, or user agents. They are stored in Vector's Supabase service as one record per install or account per day, with the latest of these totals and the tokens, tasks, and cost of each day. Vector's usage dashboard shows only aggregate totals, for example how many installs were active in a week or how many tokens were processed.

Aggregated statistics. Vector may share aggregated, non-identifying statistics drawn from usage counts, such as the number of active users, tokens processed, or popular models, with investors, partners, and the public, including through read-only links to its usage dashboard. These are totals across everyone and never include an account, install identifier, or email address. Models and custom effort levels that fewer than three accounts use are left out of anything shared this way, and so are the sessions and model use of any day on which fewer than three people were active.

Downloads. Installer downloads require a Vector account. Each download is recorded against the signed-in account with the installer platform, the version, and the time.

Retention and deletion. Usage counts, including the model-use totals, and download records are deleted automatically after about 13 months. Deleting your Vector account deletes the usage counts and download records linked to it, including earlier days reported by an install before it was linked to your account. Aggregated statistics already shared, which identify no one, are not withdrawn.

Turning it off. Usage counts are on by default. In the desktop application, turn off Settings → General → “Share usage counts” to stop check-ins, including the model-use totals. For the command-line tool, set the environment variable VECTOR_DISABLE_USAGE=1; it then sends no model-use totals, and its daily account check still runs, because sign-in depends on it, but asks Vector not to count it. Versions released before usage counts existed do not report them. To limit abuse, the check-in service keeps a keyed counter derived from the network address for one hour; it is separate from the usage counts and is not stored with them.

Website visits. Visits to public pages of the website and documentation are counted with Vercel Web Analytics, a cookie-free service that reports aggregate figures such as page views, visitors, referring sites, countries, browsers, and device types. It sets no cookies, and the address sent is reduced to the page path, without query strings. Account, sign-in, private, and shared-session pages are not counted. Vercel, which hosts the website, processes the request itself, including its network address, to produce those figures.

11. Why information is used and who receives it

We use information to provide requested features, authenticate accounts, deliver communications, answer support requests, troubleshoot failures, prevent fraud and abuse, secure the service, understand how much Vector is used through the aggregate counts in section 10, and meet legal obligations. Aggregated statistics that identify no one, described in section 10, may also be shared with investors, partners, and the public. Recipients include the account, email, hosting, payment, security, and diagnostic providers described above and the model and integration providers involved in your selected workflow.

Information may also be disclosed when necessary to comply with a lawful obligation, establish or defend legal claims, or protect people and service security. If the business is reorganized or transferred, relevant account and service records may transfer with it, subject to applicable law and appropriate notice. This Policy does not authorize unrelated use of your repository content.

Where a law requires a legal basis for processing, the applicable basis depends on the purpose: providing a service you request; legitimate interests in security, abuse prevention, service support, and measuring use through usage counts, which you can turn off, balanced against your rights; compliance with legal obligations; or consent for optional processing such as public session sharing and remote crash diagnostics. You can withdraw consent for optional processing without affecting processing that was lawful before withdrawal. Required account information is necessary to provide those features; optional diagnostics are not a condition of using Vector.

12. Storage, security, and retention

Packaged desktop builds use the operating system's secure storage for supported provider credentials and the local credential-vault key. Some application data, including local memory, uses local files instead. External runtimes have their own storage practices. Device access controls, backups, and the permissions you grant also affect security. No storage or transmission method can guarantee absolute security.

Local data remains until removed through application controls or from the relevant files, subject to local cleanup and backups. Account records are retained while needed to provide access and resolve account issues. Records of earlier purchases, and the Codium billing records described in section 4, are retained as required by law, including for tax, accounting, fraud, and dispute obligations. Payment records of those purchases are held by our payment processor, Stripe, and kept only as the law requires. Support records and diagnostic or operational logs are retained according to their troubleshooting and security purpose and applicable provider settings. Usage counts and download records are kept for about 13 months. We consider the type of record, whether the account or issue is active, legal retention requirements, and unresolved disputes when deciding retention. No single deletion period applies to every category.

Providers may process information outside your country, with different privacy laws. Locations and transfer arrangements depend on the provider and service configuration. Contact us for information about the arrangements relevant to your account. A provider's independent retention rules continue to apply to information already sent to it.

13. Your choices and privacy rights

You can choose models and integrations, limit permissions, stop background work or computer use, disconnect providers, turn off diagnostics, usage counts (Share usage counts in the desktop application, or VECTOR_DISABLE_USAGE=1 for the command-line tool), and automatic sharing, unshare public sessions, edit or erase local memory, and remove local sessions and other data. Removing a local session alone does not remove its public copy. Use the relevant provider's controls to revoke its credentials or manage records it holds. Deleting an account, deleting local data, and deleting provider-held content are separate actions.

Depending on the law that applies to you, you may have rights to access or receive a copy of personal information, correct it, request deletion, object to or restrict processing, withdraw consent, or appeal a request decision. You may also have rights relating to sale, sharing, or targeted advertising where those activities occur. These rights can have exceptions, including necessary security and legal records. We will not unlawfully discriminate against you for exercising an applicable right.

Send requests to krishnabharadwaj0521@gmail.com. Identify the account and the request without sending passwords, full card numbers, or private keys. We may need proportionate identity or authorized-agent verification before responding. We will respond within the period required by applicable law and explain a refusal where required. You can also complain to the privacy regulator where you live or work, including the UK Information Commissioner's Office where applicable. We cannot directly erase data held independently by a provider you connected.

14. Children and policy updates

Vector is intended for adults and is not directed to children. If you believe a child has supplied personal information to our account or service endpoints, contact us so we can review and address it.

We update this Policy when relevant practices change and show the effective date above. Material changes will be communicated through the service or an account contact where appropriate. We will provide any additional notice or obtain consent required by law before using information for a materially different purpose.