Privacy policy

Who and what this covers

ABot is maintained by Dor Shitrit. This notice covers ABot Runtime, its Windows Companion, and the services we operate at abot-ai.com and the ABot journal at abot.dorshitrit.com. It does not cover unrelated personal websites, independently hosted copies of ABot, or third-party services under their own policies.

For questions or requests about information we hold, email [email protected]. We cannot access or erase a self-hosted Runtime's local files merely because you use ABot.

Your Runtime data

ABot processes your messages and instructions, model responses, files and images you attach, tool inputs and results, and selected workspace information to carry out requests. Enabled features can also process saved memories, embeddings, schedules, Studio workflows and their run results.

Conversations, attachments, configuration, memories, workflow results and diagnostic logs can be saved in your Runtime environment. That environment may be your computer, WSL, a Docker volume, or another machine you administer. “Local” refers to that environment, not necessarily the device displaying the Web UI.

Diagnostic logs and model traces can contain request and response content. Credentials supplied during setup may be stored in local environment files. File permissions provide protection, but this is not a promise that all local files are encrypted at rest. Protect your account, backups and machine, and review logs before sharing them for support.

The reviewed Runtime does not automatically upload your conversation archive to an ABot-operated account service. This does not prevent a configured provider or a tool you enable from receiving content needed for its task.

What model providers receive

Requests go to the provider endpoint configured for the relevant model. Depending on the feature, this can include your prompt, earlier conversation context, system and custom instructions, retrieved memories, document excerpts, tool results and supported images. Memory embedding and SPARK learning can use a different provider from chat.

OpenAI

Using an OpenAI model sends the selected request context to OpenAI, or to the alternative endpoint you configured. Your provider credentials authenticate those requests. OpenAI's published API policy says API data is not used to train its models by default unless you opt in. That is not a guarantee of zero retention: logging, application-state retention and exceptions depend on the endpoint, account settings and provider policy. See OpenAI's API data controls. ABot does not override those terms or promise a particular retention arrangement.

Ollama

With a local Ollama server and a local model, inference can run on the machine hosting Ollama. A remote Ollama endpoint or a cloud-hosted model changes that boundary: the service processing the model request receives its context. Other enabled tools and providers may still use the network. See Ollama's local/cloud documentation and privacy policy.

Changing provider affects subsequent requests; it does not recall requests already sent. Review each provider's policies, location of processing and costs before supplying personal, confidential or third-party information.

Windows Companion and WSL / Docker

ABot AI Companion connects a Runtime to the Windows host for supported computer access and native capabilities. It is a connection component, not a hosted model provider and not a replacement for the Runtime.

The Windows Store connection flow shows the Runtime address and asks you to confirm pairing. The current invitation flow accepts a local loopback Runtime address; WSL or Docker must make the Runtime reachable through the appropriate local forwarding. The invitation contains a temporary connection grant. Treat it as a secret: do not post it in an issue, screenshot or support message.

Pairing credentials and connection state are kept on the host to support reconnection. The Runtime also receives host identity such as the computer name, operating system, OS username and home-directory path. Local status records include connection status, process identity and timestamps. These files use filesystem access controls, not a claim of application-level encryption.

A paired Runtime can send supported commands and receive computer observations or results. This can cross the boundary between a WSL/container environment and the Windows host; a container is not a guarantee that enabled host tools cannot affect Windows. The Store package is a full-trust desktop application, which is not the same as automatically receiving administrator privileges.

The Store package registers Companion to start at Windows sign-in; you can manage this in Windows startup settings. Closing its window leaves an active connection running. Companion's Disconnect action stops its connection and removes the local pairing credential, but the Runtime keeps its paired-computer record; use the Runtime's pairing controls to manage that record. Disconnecting or uninstalling does not erase data already saved by the Runtime or sent to a provider.

When native notifications are enabled, their titles, message text and navigation targets are passed to the Windows notification system. Preview text may be visible to someone looking at your screen; manage previews through your notification settings.

Microsoft handles Store distribution and updates under Microsoft's privacy statement. Legacy installation methods can involve separate downloads. Installation and Store delivery should not be confused with consent to SPARK collection or unrestricted computer actions.

SPARK and computer access are different features

Optional SPARK activity learning

When enabled, SPARK can collect application and window identity, accessible on-screen text and changes in editable controls. It processes eligible observations using the selected learning model to identify useful patterns and memories. These observations may include personal information shown by other applications.

Collection, learning and proactive suggestions have separate controls. Collection and proactive mode are off by default in the reviewed Runtime. App exclusions, activity hours and model settings let you limit processing. Redaction can reduce recognizable secrets but cannot reliably recognize every sensitive field or text value.

The passive activity collector is not a video, audio, screenshot or keystroke recorder. That distinction does not apply to separately invoked computer tools. SPARK chat can also use tools according to its permission settings; passive learning and proactive suggestion processing are not themselves permission to execute those suggestions.

Computer tools

Enabled computer tools can capture screenshots and accessibility text and perform actions such as clicking, typing and scrolling. Observations, including supported screen images, can be sent to the model selected for that request. Operating-system access and Runtime permissions still matter; some permission modes allow actions without an additional prompt for every step.

Only connect computers and applications you are entitled to access. Close sensitive content and review the active model and permission mode before a task. ABot cannot guarantee that every screenshot or accessible text fragment is free of private information.

Plugins, Studio and scheduled work

Plugins are executable code, not just descriptions in a catalog; community plugins come from third parties. They can process task input, local files, configuration, tool results and credentials provided for their services. They may send data to independent endpoints. Runtime permission declarations are not a complete operating-system sandbox.

Studio workflows and scheduled tasks can reuse saved models, instructions and permissions. Review those settings before enabling unattended work. Choosing a local chat model does not make every plugin, workflow step, embedding request or web lookup local.

Marketplace browsing and inspection involve our website and GitHub. Installing from a source repository involves the source host and the local installation flow. A Marketplace listing is not an assurance that a plugin follows this notice; inspect its source, license and own privacy disclosures.

Our website and cookies

The public product site (except Marketplace and legal pages) and journal use first-party analytics to understand page visits, section visibility, clicks and scroll depth. Events include a page path, interaction label, filtered link destination, referring host, approximate country provided by Cloudflare, device category, browser language and time. Random visitor/session identifiers are hashed before storage in the analytics database; these identifiers are pseudonymous, not guaranteed anonymous.

• abot_visitor identifies return visits and expires after approximately 180 days.
• abot_visit_session groups activity into a visit and expires after approximately 30 minutes, refreshed with accepted events.
• abot_admin is an authentication cookie for administrators, with a 12-hour session lifetime. Public browsing does not require an admin account.

The analytics client does not record what you type into form fields. Referrer and clicked-link URLs can reach the API before being reduced to a hostname or path for storage. The analytics table does not store raw IP addresses, but hosting, security and request logs may process or retain them. Cloudflare and our hosting infrastructure handle network information needed to deliver and protect the service.

The current site starts analytics automatically and does not yet provide an in-site analytics consent or opt-out control. You can block or clear cookies in your browser, but blocking cookies alone does not stop all event requests or server logging. Cookie expiry does not delete previously stored events. These privacy and terms pages do not load the site's analytics tracker.

Marketplace counters

Marketplace pages do not load the general analytics tracker. For each published plugin, a separate counter stores only its listing ID, total detail-page views and total clicks on Open in ABot. These are not unique visitor counts or completed installations. Counting begins when the feature is enabled; we do not reconstruct these totals from older visitor records.

Counter requests use no analytics cookies, persistent browser identifier or referrer. The counters database stores no IP address or hash, visitor/session ID, device information, country, browsing history or individual event record. To limit repeated requests, the API temporarily holds a keyed network identifier in bounded process memory; it is not written to the database or used to link browsing activity. Counter requests respect browser Do Not Track and Global Privacy Control signals.

Counter endpoints are excluded from application request logging and the configured Apache access logs. Network delivery and infrastructure security services can still process network information, including IP addresses. Other website requests, Marketplace submissions and existing historical analytics records have the separate handling described here; aggregate counters do not erase those records.

Comments, Marketplace submissions and support

Journal comments: we receive the text, optional display name and submission metadata. Keyed network/browser hashes help prevent abuse. Comments are held for moderation; approved comments become public. Authorized operators and the publishing agent can access pending comments. If the agent processes a comment, the text may be included in a request to its configured model provider. Do not post secrets or information you are not entitled to share.

If Cloudflare Turnstile verification is enabled on a comment form, its verification service receives the challenge token and the request IP address. This is separate from the hashed identifiers retained in the comments database.

Marketplace submissions: we retain the title, description, category, GitHub source URL, requested revision/folder, license confirmation, moderation history, inspection results and a hashed network identifier for abuse prevention. The current form does not require an email address or account. Approved listing details and repository-derived metadata become public; review notes and network identifiers are not part of the public listing. Source files remain hosted on GitHub, although we retain metadata snapshots from inspections.

Support: contacting us shares your address, message and any material you choose to attach with us and the email services delivering it. Send only what is needed. Do not send API keys, pairing links, passwords or unredacted session archives. Information posted in a public GitHub issue is public.

Storage, retention and deletion

On your Runtime: stored sessions, files, memories and configuration remain under your control. Use the available conversation, memory and SPARK controls to manage them. Logs and temporary evidence have their own limits. SPARK's pending raw observations normally have a 24-hour retention window, with cleanup while the Runtime is running or at startup; saved memories have a separate lifetime. Stopping collection is not deletion, and it does not necessarily stop processing of already collected observations.

Deleting a conversation does not automatically erase separately saved memories, workflow output files, logs, backups or copies already sent to another service. Temporary computer-tool screenshots and uploaded attachments also have different lifecycles. Provider-held copies are governed by that provider.

On our website: analytics, comments, Marketplace records and moderation/audit history currently have no fixed automatic deletion deadline configured in the application. Hiding a post or listing is not physical erasure. We use these records to operate the service, prevent abuse and preserve moderation history. Contact us to request review or deletion of information we hold; legal requirements, security needs and backup copies may affect what can be removed and when. We do not promise immediate erasure from every backup.

We use access controls and HTTPS for the public site, but no storage system or network transfer is guaranteed secure. Hosting, email, source-hosting and model providers may process information in other countries under their own arrangements. This notice does not claim a specific data residency or encryption-at-rest guarantee.

Your choices and privacy requests

You can choose model endpoints, limit tool permissions, pause optional SPARK activities, exclude applications, disconnect Companion and manage locally stored information. These controls apply to future activity or the data they specifically manage; they do not undo completed actions or third-party transmissions.

Depending on the laws that apply to you, you may have rights to access, correct, delete or obtain a copy of personal information, restrict or object to processing, withdraw consent where processing relies on consent, or complain to a privacy authority. Contact [email protected] with the relevant page, submission or context. We may need enough information to verify and locate the request; avoid sending unnecessary identity documents.

For information held only by your self-hosted Runtime, use your local controls or contact its administrator. For provider-held information, contact that provider. This policy will be updated when our practices change; material changes should be reviewed before you enable affected features.