Privacy

Model requests and saved content. Ordinary model requests pass through our gateway to the model provider. The gateway records usage and accounting information rather than prompt and response text. The web app separately stores content when you enable conversation saving, save a workspace file or create a custom web skill, as described below. Model providers receive the content needed to answer your request and have their own policies.

Conversation saving is optional. Web chat conversation saving is on by default for new accounts (from 29 September 2026; accounts created earlier keep it off unless they turn it on), and you can turn it off at any time in Settings. With saving off, the server holds an active answer temporarily so it can continue streaming after a page reload; it is not added to your saved history. Turning saving off deletes previously saved conversations. This setting does not save desktop or CLI conversations automatically, and it does not control the separate workspace files or custom web skills you choose to save.

What we store

Appearance synchronization. We store your System, Light or Dark choice against your account, its revision and update time. We also retain synchronization receipts with a change identifier, the selected appearance, before-and-after revisions and a timestamp, so retries do not repeat a change. Appearance and synchronization receipts are kept until you delete your account. This synchronization is independent of conversation saving and does not include messages, credentials, plugins, MCP connections, skills or workspace files.

Account tool catalog. Supported desktop and cloud versions synchronize skill, plugin, MCP server and connection definitions, enabled states, revisions, deletion markers and references to private skill packages. These definitions describe configuration; they do not by themselves establish a working tool connection. Catalog entries and deletion markers remain until account deletion; deletion markers prevent an older device from restoring removed entries. Retry receipts become eligible for bounded scheduled deletion after 30 days.

Cloud workspaces and runs. We store workspace names, run status and timestamps, temporary execution events, capacity leases and links to usage reservations. Events become eligible for scheduled deletion one hour after completion or the run deadline. Terminal run records become eligible after one day, except while a checkpoint, occupied capacity slot or unsettled financial reservation still depends on them. Model-selection grants record the chosen model and exact run generation and are removed with that run. Short-lived web access bindings contain hashed grants and runtime-cookie hashes, bound to your signed-in session, workspace hostname and run generation; expired bindings are cleaned up.

Files on your own computer. When Agent Cowork desktop is running, it can list what a session holds on that machine: file names, sizes, modification times and an opaque reference the desktop gives itself. Full paths stay on your computer and are never stored here. A file's contents move only when you ask for that file: the bytes are held encrypted in private Bunny Storage for up to fifteen minutes so your browser can read them, then deleted. We keep the request's state, size and checksum, not the file. Nothing is fetched in the background, and closing the desktop app stops new files being read.

Cloud project snapshots. Cloud workspaces save encrypted project snapshots in private Bunny Storage. Our database stores file paths, sizes, checksums and object references. These snapshots let a later run restore your project. Project files are separate from conversation history: withdrawing history consent does not delete project files. Replaced snapshots and deleted-account snapshots are queued for remote cleanup.

Native cloud session history. When available, this requires separate consent as well as conversation saving. It can include messages, tool results and attached images, stored encrypted in private Bunny Storage. It follows your selected conversation retention period of 30, 90 or 365 days from activity, or until deletion. Withdrawing either consent removes access to saved native history and queues remote deletion, without deleting project files. This differs from the legacy web chat described below, which does not save attached images.

Your account. The email address and name your Google or GitHub account reports, your profile image URL, and, if you signed in with GitHub, your GitHub handle. Kept while the account exists.

Sign-in tokens. The OAuth access and refresh tokens Google or GitHub issue for your account, encrypted at rest, and a session token for each browser you are signed in on. Against each session we also keep when it was created, when it was last used, and the browser and operating system its User-Agent reports, so you can recognise your own sessions and end the ones that are not yours, under Connections in your dashboard. Sessions expire 30 days after their last use; signing out, ending a session, or “sign out everywhere” deletes the row immediately.

API keys. A SHA-256 hash of each key, its first twelve characters for display, the name you gave it, and when it was last used. The key itself is shown to you once and is never stored.

Device pairing. When you pair the desktop app or the CLI, we record the app id, its version, a hash of the one-time device code, and the IP address the pairing request came from. The address is used to warn you if an approval is being requested from a different network than the one you are on, and is erased 24 hours after the pairing. The pairing record itself is kept, without the address, so that reinstalling the app retires the key it previously issued instead of one you created by hand.

Usage. One row per request: the time, the model, token counts, what it cost us and what it billed you. No prompt, no response, no file names. These rows are the accounting record and are kept for as long as we are required to keep accounting records.

Error reports. When something fails, in the web app, the desktop app, the CLI or the browser extension, we record what failed and when: the error code and our own error message, the app and its version, the session number shown in the app, and a short technical note. No prompt, no response, no file contents and no keys. The desktop app, the CLI and the extension send these reports to us in the background. They are deleted after 30 days.

Daily file transfer counters. To enforce transfer limits, we record the UTC day and total reserved bytes against your internal account identifier, plus an anonymous service-wide daily total. These counters cover private file and skill reads, including upload verification. They contain no file names, object references or content. Failed or interrupted transfers can still count toward the limit.

Image search and image generation. When your agent searches for a photograph we store the search words it sent, which library answered, how many results came back and how long it took, plus which image was chosen. Search words are text your agent composed, so treat them as you would anything else you send us. When your agent generates an image we store the plan, the size, the count and what it cost us, not the prompt, and not the image. Generated images themselves are kept in object storage and referenced by URL; they are deleted with the account.

Legacy web chat conversations, if you switched saving on. Your messages and the model's replies, the tools that ran and what they returned, and the model and mode each turn used. This is the text of your conversation: treat it as you would anything else you send us. You choose how long it is kept, 30, 90 or 365 days from the last message, or until you delete it, and you can delete any conversation, or all of them, at any time. Images you attach are sent to the model and are not saved with the conversation. This applies to the web chat only.

How each web chat turn ran. For every web chat turn we keep a diagnostics record: the model and mode used, how many steps and tools ran, timing, cost, the language, and, when images were needed, the short reason the agent gave for searching or generating them, which can quote your request. No message text otherwise. A saved conversation also keeps its working state (a short summary of the project, your stated preferences and open items), and a document the agent builds in parts keeps its latest draft under its file name. All of this is deleted when you delete your account; the working state also goes with its conversation.

Run records. Where run recording is switched on, each task also keeps a run record: which agents worked on it and in what order, each step with a short summary of what it was asked and what came back (at most 500 and 1,000 characters, with keys, passwords and card numbers removed and never a whole file), which can quote your request, the names of the files produced, timing and cost. A web chat run is deleted with its conversation, so with saving off it is gone when the turn ends, and every run is deleted with your account. The desktop app and the CLI send run records only when you switch this on in their settings.

Memory across chats. When memory is on, we keep short facts you state about yourself, your preferences, your brand, your projects and the people you work with (at most 200 entries of up to 300 characters each), so a new chat does not ask again. Memory never keeps health, money, religion, political or sexual orientation details, ID or card numbers, passwords, or anything about children. You can see, edit and delete each entry, or switch memory off, in Settings, Privacy. It is deleted when you delete your account.

Ratings you give an answer. When you press thumbs up or down under an answer, we keep the rating, the optional one-line comment you write, the model label and the time, linked to that conversation and message. We use it to improve answers. It is deleted with the conversation or your account.

Task results. Where it is switched on, each answer that made files can also keep a short result: whether the task succeeded or failed, decided by our automatic checks or by your rating, with the reason (for example, a file was missing) or your rating comment. We use it to measure and improve answers. It is deleted with the conversation or your account.

Your images and files index. A list of the images and files the agent made or found for you (name, type, size, dimensions, the prompt or search behind it, the conversation and its storage link), so you can find and reuse them. Kept until you delete the item or your account.

Payment records. A copy of each payment our payment processor reports for your account: amount, currency, fees, refunds, status and any failure reason. Card details are never part of it. Kept as an accounting record for as long as we are required to keep accounts.

Published websites. When you publish a website, we keep its link, the file it shows and whether it is public. A private link opens only for you, signed in; a public one opens for anyone who has it. Unpublishing or deleting your account removes the link.

Web pages the agent read. Public pages the agent opens are cached for 15 minutes so the same page is not fetched twice, then deleted within a day. The cache holds the page address, title and text, not who asked for it.

Private workspace files. Files you upload or ask the web agent to create are stored in private Bunny Storage. File names, types, sizes, checksums, creation times and storage references are recorded against your account in our database. Preview and download requests require an authorized account session; these files are not published through a public CDN. There is no limit on the number of files. The total and per-file sizes depend on your plan: from 100 MB in total and 5 MB per file on Free, up to 2048 MB in total and 25 MB per file on Pro, with a 512 KB limit for text files. PDF, Word, Excel and PowerPoint files are opened only inside an isolated processing service with no internet access, to read, preview or create them; its temporary copies are deleted when each job ends. Files remain until you delete them or delete your account, independently of conversation saving and conversation retention. Text the agent reads from a workspace file may be sent to the model provider to carry out your request.

Legacy custom web skills. We store the Markdown instructions you save in private Bunny Storage, with the skill name, its one-line description and aliases, enabled state, size, checksum, revision and update time in our database. You can keep up to 20 skills, 8 KB per skill and 160 KB in total. The name and one-line description of each enabled skill are included in web model requests; the full instructions are included only when a message needs that skill. Which built-in skills you switched off is stored with your account. Disabling a skill stops its use in future requests but does not delete it. Skills stay until you delete them or delete your account. They are independent of conversation saving. This legacy store is separate from the synchronized account tool catalog and its private skill packages.

Billing. Your customer and subscription identifiers at our payment processor, your plan, and your subscription status. Card details go directly to the processor and never touch our servers.

Support tickets. If you open a ticket, we store its subject and every message in it, text you wrote, which may contain whatever you chose to paste in, including code. Tickets and their messages are deleted 180 days after the ticket is resolved or closed.

Provider credentials. Operator-supplied model-provider keys are encrypted at rest with AES-256-GCM. These are ours, not yours; no customer credential is stored.

Connected services. When you connect another service in Settings > Connectors (for example Canva, Notion, Google Calendar or Google Drive), you sign in on that service's own page and it gives us an access token for your account there. We store that token, and the refresh token that renews it, encrypted, together with the service name, the account email the service reports, when you connected and when chat last used it. We never see your password for that service. Chat reads from or acts on the service only when you ask it to, and anything that creates, changes, sends, posts, pays, deletes or spends credits waits for you to approve it in the chat first. What a service returns is used to answer your request in the conversation; it is sent to the model provider for that purpose and saved only if conversation saving is on. Images, video or audio a service returns can be copied to private Bunny Storage so you can preview and download them, and are served only to your signed-in account. The service itself bills your own plan and credits with it; we do not.

Custom connectors. If you add your own MCP server, we store the name and address you entered and its token, encrypted at rest, and send your chat requests for that server to it. Remove it in Settings > Connectors to delete our copy at once; it is also deleted with your account.

Google data. Google Calendar and Google Drive connections ask only for what they need: your calendars and events (read, and change or create events you approve) and, for Drive, only the files Agent Cowork creates or you open with it. Agent Cowork's use and transfer of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements. We do not use Google data for advertising, do not sell it, do not use it to train general AI models, and people at Agent Cowork do not read it except with your permission, for security, or where the law requires. You can disconnect in Settings > Connectors or at myaccount.google.com/permissions.

How long we keep it

Appearance and synchronization receipts: until you delete your account.

Daily file transfer counters: rows older than 35 days are removed by our scheduled cleanup in bounded batches. Deleting your account immediately removes its account-linked transfer counters. Anonymous service-wide totals remain until scheduled cleanup; they do not identify your account.

Workspace files and custom web skills: until you delete them or delete your account, even when conversation saving is off. Removing access and removing the remote storage object are separate steps. We queue remote deletion and retry it if storage is unavailable. During retries, cleanup records retain an internal account identifier and object reference; the removed content is no longer available through your account. Replaced skill versions and incomplete uploads are also queued for cleanup.

Web chat conversations: not kept at all unless you switch saving on; then 30, 90 or 365 days from the last message, whichever you chose, or until you delete them. Pairing IP addresses: 24 hours. Unused device codes: one hour. Support tickets: 180 days after close. Error reports: 30 days. Sessions: 30 days from last use, or immediately on sign-out. Account data: until you delete the account. Usage and billing records: for the statutory accounting period, with the account they belong to anonymised on deletion.

Connected services: until you disconnect the service or delete your account. Disconnecting deletes our copy of the tokens at once and asks the service to revoke them. Media a service returned stays in your private storage until you delete your account.

Why we are allowed to hold it

Appearance synchronization keeps your chosen appearance consistent across supported apps. Workspace storage and custom web skills provide the files and instructions you ask us to save and use (contract).

Account, key, pairing and usage data is processed to provide the service you asked for (contract). Billing records are kept because we are required to keep them (legal obligation). Support tickets are processed to answer the question you asked (contract). We do not profile, advertise, or sell anything to anyone.

Who else sees it

Our model providers receive the content of your requests in order to answer them, and nothing about your account. Services you connect receive the requests you make of them, from our servers, on your account. Stripe processes payments and holds your billing details. Google and GitHub authenticate you. Cloudflare sits in front of the site and sees request metadata. Our application servers are in France. Bunny stores workspace files, custom skill content and generated media; its storage and delivery locations depend on the configured service. We do not sell your data.

Deleting your account

Account deletion removes your appearance and synchronization receipts, connected-service tokens and pending connector actions, account tool catalog, workspace file metadata, saved cloud snapshots and native history, and custom skill metadata. Active cloud access is revoked and running jobs are marked for cancellation; minimal run and financial fencing records can remain until cleanup is safe. Access to the associated private files and skill content ends immediately; their Bunny Storage objects are queued for deletion and retried if storage is unavailable.

Sign in, open your dashboard, and delete the account there. You will be asked to type DELETE MY ACCOUNT, and it happens immediately. It removes your email, name, profile image and GitHub handle; every OAuth token; every session; every API key; every device-pairing record; every saved web chat conversation with all of its messages; and every support ticket with all of its messages. Usage and billing rows are kept for accounting, with nothing left on them that identifies you.

If a paid subscription is still running, cancel it first: we will not cancel a live subscription as a side effect of a delete. If you would rather we did it for you, email support@agentcowork.net; we act on it within 7 days.

The desktop app

Runs on your computer. It contacts our service for sign-in, model requests and update checks. Supported versions also synchronize account appearance and the account tool catalog described above. Only explicitly published or saved project files and separately consented history participate in supported hybrid workflows; signing in does not upload every file or conversation on your computer. Legacy custom web skills remain a separate store.

Questions

Email support@agentcowork.net. If you are in the EU or UK and are not satisfied with our answer, you may complain to your national data protection authority.