Privacy Policy
Version 1.1, effective 30 September 2026.
At a glance
This summary is part of this policy and is meant to be accurate. The numbered sections have the detail.
If you have a Cosi account: we process your email address, the name you work under, the content you put into Cosi, data derived from it (tags, search embeddings, summaries), how you use the product, and error reports. We use it to run Cosi for you, to keep it secure, and to understand how the product is used. We never sell it. We do not use your content to train AI models: our language-model provider is contractually barred from training on it, and our embedding provider's standard terms permit training unless the account is opted out, which we are moving to a company account opted out of it (section 5.3).
If someone shared a treatment with you: if you only view it, rate-limit records hold your IP address and a hash of the link token for about one to two hours. Product analytics, error monitoring and font requests also occur (section 11). If you sign in to comment, we process your email address and the display name you give, and your comments are visible to the person who shared with you and to other reviewers.
If you report content to us: we process your name, email address and the notice you send, and may pass its substance to the user concerned.
If you appear in someone's images: the Cosi user who stored the image is responsible for that processing; we process it on their behalf. Section 10 tells you how to exercise your rights.
Who we share with: the providers we need to run Cosi — hosting, storage, AI, email, monitoring — listed in section 6, and authorities where the law requires.
How long: while your account exists, plus the periods in section 8.
Your rights: access, correction, deletion, restriction, portability, objection, and a complaint to the supervisory authority (section 12).
Cookies and device storage: what we need to keep you signed in, plus product analytics and error monitoring that store nothing on your device (section 9).
1. Who is responsible, and how to reach us
1.1 Controller. Cosi UG (haftungsbeschränkt), c/o Nicholas Martinelli, Bänschstraße 89, 10247 Berlin, Germany, registered at Amtsgericht Charlottenburg (Berlin), HRB 287673, is the controller for the processing described in this policy, except where section 1.2 says otherwise. Contact for all data-protection matters: hello@meetcosi.com, or by post to the address above. We have not appointed a data protection officer; requests go to the address above.
1.2 Terms used here. Capitalised terms such as Your Content, Input, Output, Actions, Viewer and Reviewer have the meaning given in the Terms of Service.
1.3 Two roles. For most of what this policy describes, we decide why and how data is processed and are the controller. For personal data inside Your Content — people who appear in the images and videos you store, people mentioned in documents you upload, and material your own clients gave you — you are the controller and we are your processor: we process that data only on your instructions, under the Data Processing Agreement that forms part of the Terms. The DPA governs personal data inside Your Content; this policy governs everything we control ourselves, including the security, abuse-prevention and notice-handling processing in section 4 that we carry out over Your Content in our own responsibility.
2. What this policy covers
The Cosi web application at app.meetcosi.com, including the public presentation, review and export pages; the Cosi browser extension; the connection to third-party services you choose to link (currently Dropbox); and our communication with you by email. It does not cover the websites you capture from or the services you connect; their own policies apply there.
3. What data we process
3.1 Account data. Your email address (used to sign you in with a one-time code), the display name and avatar you set, the declaration that you use Cosi professionally and the name you work under, your plan, and the record of which version of the Terms and this policy you accepted, when, and the wording you were shown. For a paid plan: your company or business name, optionally a VAT or tax number, and billing details (section 3.9).
3.2 Your Content. The images, videos, documents, text, treatments, decks and comments you upload, capture, import or create, together with the files you attach to assistant conversations and the prompts and search queries you type.
3.3 Capture metadata. For each asset you save with the extension, we store the web address of the file and the title of the page it came from, alongside file metadata such as filename, type and size. The extension also sends the page address and any alt text, which the ingest path does not keep in the asset record. For websites that require a login, the file address can reveal which service you were signed in to and what you were looking at. We treat the file address as part of Your Content. The website's domain and the page title also appear in the extension's usage events (section 3.7); we are removing the page title from those events.
3.4 Derived data. Data our AI features generate from Your Content to run the product: tags, search embeddings (numeric vectors used to compare images or text — we treat them as derived data, not as anonymous data), cluster names, document summaries and assistant messages. Section 5 explains how it is created.
3.5 Sharing data. The tokens behind your share links; and, for people you share with, what section 11 describes.
3.6 Connected-account data. If you connect Dropbox: your Dropbox account email and identifier, the access tokens needed to keep the connection working, and a record of the imports you run.
3.7 Usage data. Which features you use and when, technical events (for example that a save from the extension succeeded or failed, with the website's domain and the page title), performance measurements, and per-account counts of AI operations and their cost. Apart from the captured page title noted in section 3.3, usage data does not include Your Content.
3.8 Diagnostic data. Error reports from the web app, which can contain the technical state of the app at the time of an error, your account identifier and your email address. If you use the feedback button, your message and, if you choose to attach one, a screenshot are sent with the report.
3.9 Payment data (once paid plans exist). Handled by our payment provider; we receive the transaction status, customer and subscription identifiers, plan and billing information, and your billing country and location derived from your IP address for tax purposes, but not your full card details.
3.10 Support and notices. What you write to us at hello@meetcosi.com. If you send a notice about content under section 10 of the Terms: your name, email address and the notice itself; we may pass its substance to the user whose content it concerns.
3.11 Server logs. Infrastructure providers may process IP address, date and time, requested address, referring page, browser and operating system. We receive and keep no request logs from our hosting and storage provider (section 8.2).
3.12 Do you have to provide this data? Your email address, the declaration that you use Cosi professionally and the name you work under are required to create an account; without them you cannot use Cosi. Uploading content is voluntary, but Cosi cannot do anything useful without content to work on. A company or business name is required for a paid plan. A VAT identification or tax number is optional and only needed if you want to be invoiced without VAT under the reverse charge. Everything else is collected as you use the product.
4. Why we process data, and on what legal basis
4.1 Legal bases in general. Where we process data to perform the contract with you or to take steps before entering into it, the basis is Art. 6(1)(b) GDPR. Where a law requires the processing, the basis is Art. 6(1)(c). Where we rely on our legitimate interests, the basis is Art. 6(1)(f), and we have weighed our interests against yours. Where we ask for your consent, the basis is Art. 6(1)(a), and you can withdraw it at any time with effect for the future.
4.2 Purposes we control (Cosi as controller):
Purpose Data Basis Creating and running your account, signing you in, sending you sign-in codes and service messages account data, server logs contract (b) Providing the Service: storing and organising Your Content, generating derived data, running the editor, rendering exports and share links Your Content, derived data, sharing data contract (b) — for personal data inside Your Content see section 1.3 Connecting Dropbox and running imports you start connected-account data contract (b) Billing, invoicing, tax evidence and fraud prevention (once paid plans exist) account data, payment data contract (b); legal obligation (c); fraud prevention: legitimate interest (f) Keeping the Service secure: rate limiting, abuse prevention, access controls, investigating incidents server logs, usage data, account data legitimate interest (f): protecting our systems and other users Monitoring errors and reliability diagnostic data legitimate interest (f): a working product Understanding how the product is used usage data legitimate interest (f) Handling notices about illegal or infringing content, moderating, keeping records of decisions and of repeat infringers notices, the content concerned, account data legal obligation (c) under the Digital Services Act; legitimate interest (f) Answering your support requests support communications contract (b); legitimate interest (f) Evidencing the contract and defending legal claims acceptance record, notices, invoices legitimate interest (f); legal obligation (c) Emails about Cosi's own similar services to existing customers, with a free opt-out at collection and in every email email address legitimate interest (f), § 7(3) UWG Notices we must send you (security, billing, changes to the Terms) email address contract (b). These are not marketing and cannot be opted out of
4.3 Processing on your instructions (Cosi as processor). Personal data inside Your Content is processed only to provide the features you use, on your instructions, under the DPA. We name no legal basis for it here; the basis is yours to hold as controller.
4.4 No sale, no advertising, no re-identification. We do not sell personal data and do not share it with advertisers. We use no advertising trackers. Where we use data in de-identified or aggregated form, we do not attempt to re-identify it.
5. AI processing
5.1 What happens, feature by feature.
When you upload or capture an image, a reduced-size copy is sent to our tagging provider, which returns descriptive tags, and the image together with those tags is sent to our embedding provider, which returns a numeric vector used for similarity search. Videos are not analysed by AI today; only a still frame is generated for display.
When you search, your query is sent to our language-model provider to interpret it, and to our embedding provider to compare it against your assets.
When you upload a document such as a brief, it is sent to our language-model provider, which returns a structured summary. We store the summary.
When you use the assistant, your messages, the parts of your project the assistant needs, any images or files you attach, and — on request — the pixels of an image or a rendering of your current page are sent to our language-model provider, which returns the reply or the edit.
When groups of similar images are named, a sample of their tags is sent to the language-model provider.
When you export a deck, it is rendered in a browser on our job-processing provider's infrastructure.
5.2 Providers. The providers are listed in section 6. We promise capabilities, not models; if we change a model or provider, we update this policy and tell you under section 16.
5.3 Training and retention. Our language-model provider (Anthropic) is contractually barred from training on your Input and Output. It deletes API inputs and outputs within 30 days of receipt or generation, keeps content flagged for a usage-policy violation for up to two years and the related trust-and-safety classification scores for up to seven years, and its safety systems may decline a request. Our tagging provider (OpenRouter) carries tagging requests under a contractual zero-data-retention commitment: it deletes inputs and outputs promptly after the reply is generated, and with zero-retention routing switched on it will not send a request to a host that is not zero-retention — it refuses the request instead. We pin those requests to a single host, DeepInfra, whose terms state that it does not store, sell or train on the data without explicit consent. Our embedding provider (Voyage AI) applies standard terms that permit it to use content to train and improve its models unless the account is opted out; our account is opted out, and on an opted-out account content is deleted immediately after processing. We will tell you before any of this changes.
5.4 Your rights are with us. You exercise your data-protection rights with Cosi, not with our AI providers. Their own privacy policies do not apply to what they process for us on your behalf.
5.5 People in images. We do not identify people in images, do not use facial recognition, and do not ask our models to infer sensitive characteristics such as ethnic origin, religion, health or sexual orientation. Automated tagging describes what is visible, including people's apparent gender, approximate age and appearance, so that you can find images in search. Tags may also note which director's or photographer's visual style an image evokes; that names a style, not a person in the image.
6. Who receives data
6.1 Processors. The following providers process data on our behalf under data-processing agreements. Annex 2 of the DPA uses the same identities, locations and safeguards for those processing data inside Your Content on our instructions (DPF = EU–US Data Privacy Framework; SCCs = Standard Contractual Clauses).
Provider Role Location Transfer safeguard Anthropic, PBC language models: the assistant, document summaries, search interpretation and cluster naming. Which model runs a feature is our choice and may change (Terms section 3.3) USA; API inputs and outputs deleted within 30 days (section 5.3) SCCs Voyage AI Innovations, Inc. (a MongoDB company) image and text embeddings for search USA; our account is opted out of training (section 5.3) SCCs OpenRouter, Inc. routes automatic image tagging to the model host under zero-data-retention settings; also carries the fallback language model for tagging USA; retains no prompts (section 5.3) SCCs DeepInfra, Inc. runs the open-weight vision model that produces image tags; the only host we allow tagging requests to reach. It is OpenRouter's sub-processor, not a provider we contract with directly; we name it because it sees your images USA; does not store or train on the data (section 5.3) Through OpenRouter's SCCs Supabase, Inc. authentication, database Data stored in Ireland (eu-west-1); vendor support may access from outside the EU SCCs Cloudflare, Inc. application hosting, content delivery, asset storage (R2, EU jurisdiction); we receive no request logs from it Data at rest in the EU jurisdiction; global edge network SCCs Resend, Inc. sending sign-in codes and service emails; email content and delivery logs kept for 30 days USA SCCs Trigger.dev, Inc. background processing: image and video processing, AI processing, PDF and PowerPoint export Jobs run in Germany (eu-central-1); the vendor is US-based and support may access from outside the EU SCCs PostHog, Inc. product analytics in the web app and the extension; identifies you by your account identifier and email address; IP addresses anonymised; events kept for one year; session replay off EU None needed (EU) Functional Software, Inc. (Sentry) error monitoring and the feedback widget in the web app; error context includes your account identifier and email; events kept for 30 days EU (de.sentry.io) None needed (EU)
6.1a Payment provider (once paid plans exist). Stripe Payments Europe, Ltd. (Ireland) processes payment details as an independent controller under its own privacy policy, with transfers to Stripe, Inc. (USA) under the DPF. We receive the transaction status, customer and subscription identifiers, plan and billing information, and the billing country, but not full card details.
6.2 Fonts. Our pages load fonts from Google Fonts, which means Google LLC (USA) receives your IP address when a page loads, including on shared presentation and review pages. Basis: legitimate interest (f) in a consistent typeface. We are moving to serving these fonts from our own servers, after which no font request will leave our infrastructure.
6.3 Categories of other recipients. Professional advisers (lawyers, accountants, insurers) where necessary for their services to us; internal tools for support, debugging and development — error monitoring, issue tracker, team chat, and AI coding tools used in development — which can contain Your Content while we support you or investigate a problem; a successor in the event of a merger, acquisition or sale of our business, subject to this policy; and authorities, courts or other third parties where the law requires or permits disclosure, in particular to comply with a legal obligation, to respond to a lawful request, to protect the rights, safety or property of Cosi, our users or others, or to enforce the Terms.
6.4 Connected services. Dropbox is not our processor. When you connect it, or use the extension on it, Dropbox processes your data as an independent controller under its own policy.
6.5 Changes to processors. We tell you by email at least 14 days before we add or replace a processor we engage directly that handles Your Content. Where one of our processors changes its own sub-processors, we pass on the notice we receive without undue delay.
7. International transfers
Some of our processors are in the United States or process data there: our AI providers (Anthropic, Voyage AI, OpenRouter and DeepInfra), our email service (Resend) and our job-processing provider (Trigger.dev), whose jobs now run in its German region although the vendor itself is US-based. Cloudflare keeps your files in its EU jurisdiction but runs a global edge network. Where the European Commission has issued an adequacy decision, including for companies certified under the EU–US Data Privacy Framework, we rely on it. Otherwise we rely on the Standard Contractual Clauses under Art. 46(2)(c) GDPR, supplemented where needed. The table in section 6 names the safeguard per provider. You can ask us for a copy of the safeguards at hello@meetcosi.com.
8. How long we keep data
8.1 In general. We delete personal data as soon as it is no longer needed for the purpose it was collected for and no legal retention duty applies. Where we must keep data for a legal reason, we restrict it: it is blocked and not used for any other purpose until it can be deleted.
8.2 Periods.
Data Kept Account data and Your Content while your account exists; deletion starts on request without a restore period and completes asynchronously (section 8.3) Projects you delete 30 days in Recently Deleted, then the project records and cascading child rows are removed by the next daily purge; the archive assets used in them are not deleted by that purge Archive assets you delete until you delete your account — deleted assets are hidden but not yet purged Derived data (tags, embeddings, summaries, assistant conversations) follows the asset, project and account it belongs to Backups up to 7 days after deletion from live systems (daily database backups, no point-in-time recovery); file storage keeps no object versions, so once a file is deleted from storage no copy is retained Connected-account tokens until you disconnect or delete your account Comments you leave as a reviewer until deleted by the author, removed with the relevant asset or project record, or through account deletion Server logs we receive and keep no request logs from our hosting and storage provider Rate-limiting records (IP address, link-token hash, request scope, window timestamp and count) normally about one to two hours Analytics events one year (IP addresses anonymised) Error reports 30 days Support emails for as long as needed to deal with the matter and to answer follow-up questions Emails we send you (sign-in codes, service emails) our email service keeps content and delivery logs for 30 days Terms acceptance record, content notices and our decisions on them 3 years after the account closes (limitation period, §§ 195 ff. BGB) Invoices, accounting records and business correspondence for the periods German tax and commercial law require Inactive free accounts deleted after 12 months without sign-in, with 3 months' notice by email
8.3 Deleting your account. When you delete your account, deletion from our live systems starts immediately; there is no restore period. The deletion runs as a queued job with retries: it removes your files, then your records, then your sign-in identity, and completes asynchronously. It does not yet remove your profile from our analytics provider (kept up to one year), our error-monitoring provider (30 days) or the email logs at our email service (30 days); those follow the vendors' own retention. Records we must keep, and copies in backups, follow the table above.
8.4 Erasure requests. If you ask us to erase your data under Art. 17 GDPR, we do not wait out the Recently Deleted period; we delete the data directly, subject to the legal retention duties above.
9. Cookies and storage on your device
9.1 Strictly necessary. To keep you signed in and to run the app we store your sign-in session, editor settings such as open tabs and view preferences, temporary connection data for Dropbox, and temporary files for downloads in your browser's local storage. This is necessary to provide the service you asked for (§ 25(2) no. 2 TDDDG) and needs no consent.
9.2 Analytics and error monitoring. We use PostHog to understand how the product is used. It stores nothing on your device — no cookie, local storage or session storage — so repeated visits are not recognised from your browser, including on shared presentation and review pages. Events are linked to your account while you are signed in. IP addresses are anonymised and session replay is switched off. Our error monitoring (Sentry) likewise stores nothing on your device; its session replay is switched off. We are adding a switch in Settings to turn usage reporting off; until then, you can object at any time under section 12.2 and we will stop. Basis: legitimate interest (f).
9.3 No advertising trackers. We use no third-party advertising or tracking cookies.
9.4 Extension. The extension keeps your sign-in token and your project list in the browser's extension storage so that it can save to your archive. It sends the technical events described in section 3.7, associated with your account identifier when you are signed in. The extension's analytics client does not pass your email address in its event properties.
10. People who appear in your content
If you appear in an image, video or document that a Cosi user stored, that user is the controller for that processing and is responsible for having a legal basis and for any information owed to you. We process the material on their behalf and do not use it to identify you (section 5.5). To exercise your rights, contact the person who shared the material with you or who stored it. If you do not know who that is, write to hello@meetcosi.com: we will pass your request to the user, and we act ourselves where we have specific knowledge that processing is unlawful.
11. If someone shared a treatment with you
11.1 Viewing. You need no account to view a presentation or review link. For rate limiting, the review backend stores a key containing a hash of the link token and your IP address, with a time window and request count; normal cleanup removes it after about one to two hours. We keep no view counter and no log of who viewed a link, and the owner is not given view analytics. Shared pages also run our product analytics (section 9.2), our error monitoring (section 3.8) and load fonts from Google (section 6.2). The page address sent to our analytics contains the link token; we are removing that.
11.2 Commenting and flagging. To comment or flag on a review link, you sign in with your email address and give a display name. We then hold an account for you under section 3.1. Your display name, your comments and the time you left them are visible to the person who shared with you and to other reviewers of the same project. Cosi is not responsible for what the owner does with your comments outside Cosi.
11.3 Removing a comment. You can delete your own comments in the review view while you have access. Soft-deleting the project removes access; comments are deleted when the project is permanently purged or through account deletion. Contact us if your access has ended.
11.4 Onward sharing. The owner copies and distributes share links and may share them with others; anyone with the link can view. Cosi has no recipient-invitation email flow. The only email from Cosi that contains a link to a shared page is the sign-in email we send when you choose to sign in from a review link; it returns you to that page.
12. Your rights
12.1 You have the right to access the personal data we hold about you (Art. 15 GDPR), to have it corrected (Art. 16), to have it erased (Art. 17), to restrict its processing (Art. 18), and to receive the data you provided to us in a portable format (Art. 20). Where processing is based on consent, you can withdraw it at any time with effect for the future.
12.2 Right to object. Where we process your data on the basis of our legitimate interests (Art. 6(1)(f) GDPR), you have the right to object at any time, on grounds relating to your particular situation (Art. 21(1)). We then stop the processing unless we can show compelling legitimate grounds that override your interests, rights and freedoms, or the processing serves the establishment, exercise or defence of legal claims. Where we process your data for direct marketing, you can object at any time without giving reasons, and we stop (Art. 21(2)).
12.3 How to exercise your rights. Write to hello@meetcosi.com. We answer within one month; for complex requests we may extend by up to two further months and tell you why. We may ask you to verify your identity. On request we return by email, within 30 days, your original files, asset metadata and tags, deck data, documents, assistant conversations, preferences, your own comments (other people's redacted) and your usage log; this also applies after a paid plan has ended. Self-serve, you can export a deck as a PDF or as an editable PowerPoint file, download single assets and download a project's assets as a ZIP; no export reconstructs a project, because a project's layout, assistant state, preferences and comments are not in any file you can download today. Nothing is retrievable after account deletion. Some limits apply: we cannot give you a copy of data that would reveal another user's content — a reviewer cannot obtain the owner's treatment through an access request — and we cannot delete copies that viewers downloaded before a link was revoked.
12.4 Complaint. You can complain to a supervisory authority, in particular in the member state of your residence or of the alleged infringement. The authority responsible for us is the Berliner Beauftragte für Datenschutz und Informationsfreiheit, Alt-Moabit 59–61, 10555 Berlin, mailbox@datenschutz-berlin.de.
13. Automated decision-making
We make no decisions about you based solely on automated processing that produce legal or similarly significant effects (Art. 22 GDPR). AI tagging, search and the assistant help you organise and build; they do not make such decisions about you.
14. Security and operational access
14.1 Measures. We protect Your Content with access controls that isolate each user's data at the database level, enforced by automated checks in our development process; private storage served through short-lived signed addresses (presentation media for at most one hour; other addresses typically one hour and at most 24 hours); encryption in transit and at rest provided by our infrastructure providers, at provider level and without per-tenant keys; one-time-code sign-in; and secrets kept in a managed vault. Share links use long random tokens designed to resist guessing; anyone holding the link can use it, and they do not restrict viewing to a named recipient.
14.2 Who at Cosi can see Your Content. Administrators at Cosi can access Your Content for support and operations, under the confidentiality obligation in the Terms (section 4.9) and only to the extent needed. We do not browse users' archives.
14.3 If something goes wrong. If a personal-data breach is likely to result in a high risk to you, we tell you without undue delay. If you believe your data has been misused, lost or accessed without authorisation, tell us at hello@meetcosi.com.
15. Children
Cosi is for adults. We do not knowingly offer accounts to anyone under 18; if we learn an account holder is under 18, we close the account and handle deletion under section 8. Images or other content may depict or mention minors; that processing follows sections 1.3 and 10.
16. Changes to this policy
Each version carries a date. We tell you by email about material changes, and about any change to the processors that handle Your Content (section 6.5), before they take effect. We keep earlier versions and send you the version that applied at any given time on request.