This is a courtesy translation. In case of doubt, the German version prevails.
This policy applies to Nema Crew with the apps for Mac, Windows, iPhone and Android, the website crew.nema.im, the programming interfaces (API) of the Nema Services and the technical addresses under workspace.nema.im, which today only redirect to crew.nema.im or complete sign-ins, invitations and payments. A separate privacy policy at nema.im/legal/privacy applies to the website nema.im.
1. Controller
Leonardo Antonio Carta (sole proprietor)
c/o Emperon Studio GmbH
Sachsenring 51a
50677 Cologne, Germany
Email: support@nema.im
Phone: +49 221 27645575
We have not appointed a data protection officer. The contact person for data protection is Leonardo Antonio Carta at the email address above.
2. Who is responsible for which data
That depends on how you use the Nema Services:
- As a private individual or in your personal area, we are responsible for the processing of your data. This policy describes what we do.
- In an organisation that a business uses for its work, the organisation is responsible for the content of that organisation (channels, agents, emails, data from connected services). We process this content only on its behalf and according to its instructions (Data Processing Agreement, crew.nema.im/legal/dpa). Please contact the organisation first with questions about this content.
- We are in any case responsible for your account, sign-in, billing and the security of our systems.
- If another provider operates a module, it may be responsible for that processing itself (Section 11.3).
- If you use the Claude Code or Codex runtimes, Anthropic or OpenAI process your requests under their own responsibility (Section 6.4).
3. Account and registration
Account: We process your name, email address, your password (only as a bcrypt hash, not in plain text), language setting, time of registration and your roles in organisations. Optionally, you can state a form of address (male, female, non-binary); it is used only so that the agents address you correctly and is sent with the task to the language model for that purpose (Section 6). Without it, the agents use neutral wording. You can change or remove it at any time in the settings.
Minimum age: At registration, you confirm that you are at least 18 years old; we store the time.
Consent to the Terms: At registration, we store which version of the Terms you agreed to, with time, language, IP address and browser or app identifier (user agent). We merely point you to this Privacy Policy at that point; it does not require consent (Section 21).
Confirmation of the email address: After registration, we send you a six-digit code and a link. Both are valid for 24 hours. We store them only as a check value (hash), count failed attempts and sends to prevent misuse, and store the time of confirmation. Until confirmation, you cannot use the Nema Services: you can only confirm, change the address, manage your profile and password, retrieve the legal texts, export and delete your account. Until then we do not create any agents or workshop for you. If the address of an account is not confirmed 30 days after registration, we delete the account automatically with all associated data.
Purpose: Provision of the account, sign-in, proof of consent.
Legal basis: Art. 6(1)(b) GDPR; for proof of consent, Art. 6(1)(f) GDPR (our legitimate interest in being able to prove the conclusion of the contract and the applicable terms).
3.1 Two-factor sign-in and passkeys
Both are optional; you turn them on and off in the settings. Checks take place in the sign-in service on our own server in the EU (Section 20). Nothing goes to third parties for this, and we do not use any further service providers for it.
Two-factor sign-in with an authenticator app: After your password, we ask for a six-digit code from an authenticator app of your choice (TOTP method under RFC 6238). We store:
- the secret key from which the app and the server calculate the codes, encrypted (AES-GCM); during setup, until the first correct code, likewise encrypted, for at most 15 minutes,
- the time it was turned on,
- the last accepted time step, so that the same code works only once,
- 10 recovery codes in case you do not have your app to hand, only as a check value (SHA-256 hash), and when a code was used. We show you the codes themselves only once. If you have new ones generated, they replace the previous ones.
If you turn two-factor sign-in off, we delete the key, the time it was turned on, the time step and all recovery codes.
Protection against trying out codes: After 10 wrong codes within 15 minutes, we accept no further attempts for your account for 15 minutes. For this, we count failed attempts per account under a checksum in a cache; the entries expire after 16 minutes at the latest. In addition, we limit attempts per IP address, likewise only under a checksum and for one minute. One-time values of the sign-in (after the correct password, an interim key for the second step, the task for the passkey check, the code for returning to the app) are likewise kept only under a checksum, for at most 5 minutes, and deleted when first used.
Passkeys: A passkey serves as a second factor or for signing in without any password (WebAuthn/FIDO2 method). Your device or your password manager creates a key pair for this. The private key never leaves your device or your password manager; we receive only the public key. Whether and where your device or password manager synchronises the passkey is determined by its provider. For each passkey we store:
- its identifier (credential ID) and the public key,
- the signature counter, by which we recognise a copied key,
- the transports your device reports (such as USB, NFC, Bluetooth or built into the device),
- the type identifier of the authenticator (AAGUID); it denotes the model of the security key or the password manager, not you as a person,
- whether your device reports that the passkey is backed up or synchronised,
- the name you give the passkey,
- the times it was created and last used.
Creating a passkey and signing in with one take place on a page at api.nema.im (for the domain nema.im) in your browser or in a system window of your device. The page loads nothing from third parties; accessing it creates the server logs under Section 13.1. You return to the app with a one-time code that only the app which started the process can redeem (PKCE method). If you remove a passkey in the settings, we delete its entry immediately; you remove it from your device or password manager yourself.
Sign out other devices: If you choose this when turning on two-factor sign-in, we store the time. Sign-ins that began before then are no longer valid afterwards.
Notice by email: If you turn two-factor sign-in on or off or create a passkey, we send an email to your account's address stating the type and time of the change, for a passkey with the name you gave it, via our mail server (Section 12). This lets you notice if someone else changes how your account signs in.
Purpose: Protecting your account against access by others; sign-in without a password.
Legal basis: Art. 6(1)(b) GDPR for two-factor sign-in, passkeys and signing out other devices, which you turn on yourself; Art. 6(1)(f) GDPR for limiting failed attempts and for the notice emails (our legitimate interest in protecting accounts against takeover and informing you about changes to your sign-in); in addition Art. 32 GDPR.
Retention: until you turn two-factor sign-in off or remove the passkey, at most for as long as the account exists; recovery codes until new ones are generated; the time of “sign out other devices” until the account is deleted. If you delete your account, we delete all of this immediately (on backups, Section 17.6).
3a. Student plan: verification by university email
If you apply for the Student plan (Terms Section 4.1a), we process:
- the university address you enter and its domain,
- the time of the application and thus of your declaration in the checkbox that you study, do a PhD or teach and use Crew only for studying, research or teaching,
- the link we send to the university address, only as a check value (hash); it is valid for 24 hours,
- the times of confirmation, expiry and, where applicable, ending together with its reason, which reminders we have sent and how many links were sent per day.
We check the domain only against our own list and patterns of university domains. We use no external verification service, do not ask your university and do not pass the details on. We send the email with the link via the same mail server as our other emails (Section 12) to the university address; like any email to this address, it reaches your university's mail server. Reminders before expiry and the notice on expiry go to the address of your account.
Purpose: proof that you may use the Student plan, and protection against misuse, for example that a university address applies to only one account at a time and that a blocked account does not apply again. Legal basis: Art. 6(1)(b) GDPR for checking and providing the plan; Art. 6(1)(f) GDPR for protection against misuse and proof of your declaration (our legitimate interest in limiting the free plan to eligible persons and being able to prove this). The details are voluntary; without them you cannot use the Student plan. Retention: as long as the verification is valid. 12 months after it ends, whether it expired, was ended by you or was never confirmed, we delete the details completely. If we ended the plan because of false details or misuse, we delete the university address and the domain after 12 months; only the block, with its times and reason, then remains stored on your account until we lift it or the account is deleted. If you delete your account, we delete all these details immediately (for backups see Section 17.6).
4. Content in the Nema Services
We store what you and your agents write and generate: chats, messages in channels, direct messages and threads, instructions and settings of your agents, memories and knowledge entries, uploaded files and attachments (in our database), previews and the steps of an agent run (which tool was called with which result).
For an agent's instructions, we also keep the earlier versions, whether a person saved them or the agent itself (Terms Section 3.2). We store them with the time, the author and, for changes by the agent, its short reason, so that changes can be traced and undone. Proposals by an agent that you reject are included. We delete these versions together with the agent.
Files for your devices: An agent can send you a file from its workshop so that you can save it on your Mac, Windows PC, iPhone or Android device. To do so, we copy the file as an attachment into our database and show it in the conversation as a card. In a channel with other people besides you, it appears in your direct message with the agent instead. Only you can download it with your login. Your apps also receive a signed link with which the file can be retrieved for about 7 days without logging in; do not pass this link on. When one of your devices saves the file, we record on the message the device type, the device name (such as “Anna’s iPhone”), an identifier of the device and the time; we do not learn where on the device the file is stored. Other agents in the same channel or thread see the file like other attachments and may receive it as input for their next reply. The file remains stored until you delete the message, the file or your account (Section 17) and, unlike the workshop, is included in our backups (Section 17.6).
For searching memories and knowledge and for selecting the answering agent in channels (Section 5.1), we create numerical vectors from your content. This is done on our own servers with a locally operated model; nothing goes to third parties for this.
Purpose: Provision of the service, in particular so that your agents remember earlier conversations and decisions.
Legal basis: Art. 6(1)(b) GDPR; in organisations we process the content on their behalf (Section 2).
Content often contains personal data of third parties, such as names of customers in an email. We process this data on your behalf or, in the personal area, on the basis of our legitimate interest in providing you with the service you requested under Art. 6(1)(f) GDPR. You check whether you may entrust this data to the Nema Services.
5. How agents work
Agents are triggered by a schedule, events from connected services (webhooks), regular queries of these services, incoming emails or a message from a human or another agent. In each run, we process the triggering event, the history needed for the task and the results of the tools called.
We store the steps of a run for 7 days in a cache so that you can follow what an agent did. Up to 50 steps of a run are additionally stored with the agent's response; they and the other messages an agent writes remain stored like other content (Section 17).
If an agent of the Nema runtime searches the web, the search term goes from our server to a search engine; your IP address is not transmitted. If an agent retrieves a website, its operator receives the request from our server.
Agents make no decisions within the meaning of Art. 22 GDPR about you that produce legal effects concerning you. Whether you yourself use agents for decisions about others is your responsibility (Terms Annex 2).
Emails that an agent sends via one of its mailboxes always carry a notice that they were created with the help of AI (Terms Section 3.8).
5.1 Who answers in a channel
If a human writes in a channel with several agents without @-mentioning an agent, we do not wake all agents of the channel but at most two that fit best. If none of them answers, we wake the next best once. Nothing changes with an @-mention, in direct messages or between agents. The selection only determines which agent answers; nothing is decided about you or other persons (Section 5).
We check in turn:
- in a thread: which agent is working there at the moment or answered last;
- otherwise the language model of the Nema runtime decides as an overarching router, even if all agents of the channel work with Claude Code or Codex. For this we send the new message, up to eight previous messages of the same channel or thread (shortened) and short descriptions of the agents in question with their names to this model, by the same route and to the same recipients as for the Nema runtime (Section 6.1). Humans appear in it only as “Human”. If the organisation has switched pseudonymisation on, it replaces names and other details within the texts to the extent the organisation has set and with its limits (Section 7.1); if it is switched on but not possible at the moment, this step does not take place. If it is switched off, or in personal spaces, the texts go to the model unchanged, as with the Nema runtime itself. The model only answers which agents should answer;
- only if step 2 does not decide (model not reachable, time limit, pseudonymisation not possible at the moment): whether the message names an agent, otherwise the agent who wrote there last, otherwise the one whose description (name, job, when to ask it) fits the message best. For this comparison, and to order the agents for step 2, we create numerical vectors with a model we run on our own server; nothing goes to third parties for this. We do not store the vectors of the messages; we keep the vectors of the descriptions in memory only.
Of the selection we store only which agents were woken, which were not and who steps in, without content, for up to 24 hours in a cache. Our logs of it contain no content (Section 13.1).
Purpose: that the right agent answers, instead of all agents of the channel processing the message and sending it to their language model.
Legal basis: Art. 6(1)(b) GDPR; Art. 25 GDPR; in organisations on behalf of the organisation (Section 2).
6. Runtimes: where your content goes for processing
For a language model to respond, it must read the request. What is sent is the conversation context, not just the last message: what is at the beginning of a conversation is sent again with further requests, as are results of tools and content from connected services, insofar as the task needs them.
6.1 Nema runtime
The request goes to an intermediary service for model interfaces and from there to the operator of a language model. These are two recipients. We ourselves specify which operators may handle a request: with every request we give the intermediary service a fixed list of operators for the respective model, prohibit it from falling back to other operators, and allow only operators that the intermediary service lists as not storing requests (zero data retention). If none of the specified operators is available, the request fails instead of going to another operator. Who the operators are and where they process is stated in the list of providers (Section 15).
Images that you or others send to an agent in a channel or a direct message are seen by the language model itself: they go with the request to the operator of the language model. Other images and videos that are to be viewed go the same way to the operator of a model for image analysis. This includes screenshots of your computer and of an agent’s own screen (Section 8a) as well as images in results of modules (Section 11).
If your organisation has switched on pseudonymisation, details are replaced before sending, in images too if the organisation so chooses (Section 7). Otherwise the content is transmitted unchanged.
Legal basis: Art. 6(1)(b) GDPR. The intermediary service and operators process the content on our behalf (Art. 28 GDPR): with the intermediary service, its data processing agreement applies, which is part of its terms of service; it uses the operators as its sub-processors. On the transfer to the USA see Section 16.
6.2 Claude Code runtime with your own Claude subscription
You sign in with your own account at Anthropic. The sign-in token is stored in a container on our server assigned only to you; in our database we store only whether the connection exists, the email address and the type of subscription. Model calls go via a proxy of our own on our server to Anthropic (USA). If pseudonymisation is switched on, terms are replaced on the way (Section 7). The sign-in itself takes place directly between you and Anthropic.
6.3 Codex runtime with your own ChatGPT subscription
You sign in with your own account at OpenAI (sign-in via a device code). The credentials are stored in a container on our server assigned only to you. Model calls go via a proxy of our own on our server to OpenAI (USA). The following do not run via this proxy and therefore without pseudonymisation: sign-in and renewal of the credentials, and auxiliary calls by Codex to OpenAI (for example regarding account and settings).
6.4 What applies to both own subscriptions
Anthropic and OpenAI are not our processors here. They process your requests on the basis of your own contract with them and under their own responsibility; whether they store your requests or use them for their models is governed by that contract and your settings there. We transmit the content because you have chosen this runtime.
In organisations, these runtimes are available only once an Admin or Owner has confirmed once that the organisation bears responsibility for the processing by the respective provider. The organisation thereby decides, as controller, that content of the organisation may go to these providers under a member's own subscription; we transmit it on this instruction. We store who gave the confirmation and when. If it is withdrawn, agents with these runtimes no longer start a new run in the organisation. We do not check which account a member uses to sign in with the provider.
Legal basis: Art. 6(1)(b) GDPR (execution of the runtime you have chosen); in organisations, the organisation’s instruction under Art. 28(3)(a) and Art. 29 GDPR.
7. Pseudonymisation and field masks
7.1 Pseudonymisation
Admins and Owners of an organisation can switch on pseudonymisation. The Nema Services then replace personal and confidential details in the messages of a conversation that go to a language model with placeholders, such as PER_k3x9, and translate the response back. Pseudonymisation is switched off until the organisation switches it on, and does not apply to personal areas without an organisation. It works by the same rules in all three runtimes (Nema, Claude Code and Codex).
The organisation decides in its privacy settings what it covers. The following is replaced:
- the terms on a list maintained by the organisation itself, such as customers or products;
- if the organisation switches on automatic detection, what a detection model that we run ourselves on our server finds: names of persons, companies and places, email addresses, phone numbers, IBANs, ID and other reference numbers, dates of birth and health information (each of these types can be deselected individually);
- if the organisation also switches on images, also in images, screenshots and PDF files that go to a model, that is, in uploaded images and attachments, screenshots of your computer and images in tool results. We read the text in them and make recognised details unrecognisable, in images by pixelating, in PDF files by covering, each with the placeholder on top. PDF files then go to the model as an image PDF. Videos cannot be processed this way; if images are switched on, videos therefore do not go to the model for image analysis.
The organisation also decides whether results and events of a module and results of your own MCP server go to the model pseudonymised (default: yes, per module and per server), and for your own MCP servers also whether they receive the placeholders instead of the real details (Section 11.5).
Detection runs exclusively on our server; nothing goes to third parties. We do not store the original text additionally for this; it is kept only where it is stored anyway, such as in the history (Section 4). To speed things up we keep uploaded images and files in memory for at most 15 minutes.
What we do store is the result for each section of text: where in the section something was found (start and end) and of what kind it is, such as person or IBAN. These results never contain the text or the term found. The key under which a result is kept is a checksum (HMAC) of the section, together with the organisation's terms, formed with a secret key. Without this key you cannot work out from a known text whether a result exists for it, and the result carries no account, organisation or conversation identifier. We store it so that a long history does not have to be recognised again with every call, after every update of our services and in every new session: in memory, in a cache (up to 14 days) and in our database. We delete a result in the database when it has not been used for 60 days; it is included in the backups (Section 17.6). Health data (Art. 9 GDPR) may be recognised in the process, solely to hide it from the language model. In organisations this happens on behalf of the organisation (Data Processing Agreement).
We also store the finished result: the text in which the detected details have already been replaced by placeholders, with the position and type of the placeholders, so that the same text does not have to be processed again. The detected terms are not in it. The result belongs to a conversation, for Claude Code and Codex to the member in the organisation, and is kept under a checksum (HMAC) of the original text formed with a secret key. We delete it with the conversation or the organisation and otherwise when it has not been used for 45 days; it is included in the backups (Section 17.6).
So that the back-translation works, we store a mapping of placeholder to term: in the Nema runtime per conversation, for Claude Code and Codex per member and organisation. In addition, we keep this mapping per organisation so that a placeholder can also be translated back outside the conversation in which it was created. The terms in it are encrypted. The mapping per conversation is deleted together with the conversation, the account or the organisation; for Claude Code and Codex it exists until the account or the organisation is deleted. The mapping per organisation exists until the organisation is deleted, even if the conversation or the account an entry comes from has already been deleted. We store the list of terms maintained by the organisation itself unencrypted.
If pseudonymisation is switched on but technically not possible at the moment, the request is not sent. This applies in all three runtimes, including when memories are derived from conversations.
Where it does not apply. You should know this before relying on it:
- Detection is not complete. In our tests about 98 % of sensitive spans were replaced; paraphrased health information and texts in Spanish, Portuguese or Turkish are recognised less often. Whatever is not recognised goes out unchanged.
- Without automatic detection, only what is on the list is replaced; images and PDF files are processed only with images switched on, otherwise they go to the model unchanged.
- In images, only text is recognised. Faces and other image content are not made unrecognisable.
- Input to modules, your own MCP servers and other tools outside the Nema Services goes out in plain text, because the tool is supposed to work with the real terms. Exception: your own MCP servers for which the organisation has switched on “Send placeholders” (Section 11.5). Web search terms go to the search engine with the real terms.
- Codex: see Section 6.3.
- A language model can infer a person from the context even if their name has been replaced. Pseudonymised data remains personal data.
7.1a Sharing examples to improve detection
You can send us examples where our detection (Section 7.1) got something wrong. This is voluntary and happens per example:
- When you report that a term is not personal data, there is a box “Share this example to improve detection”. It is off by default. Without the tick, all we store from the report is the entry in the organisation's exception list (Section 7.1), no example.
- When you report that something should have been detected, the report itself is the sharing. The dialog tells you beforehand what will be reported. If you do not want to share anything, cancel.
Only in organisations: You can share examples only in an organisation, because only there is pseudonymisation available (Section 7.1). This is possible only if an Owner or Admin has switched on “Members may share examples to improve detection” in the organisation's privacy settings. The switch is off by default and can be switched on no earlier than 5 December 2026 (Data Processing Agreement, Section 3a). We store who switched it on or off and when. While it is off, the organisation has neither the box nor sharing via “should have been detected”. When it is on, you still decide yourself per example: the switch allows sharing, it shares nothing.
What we store:
- the term and its type (for example person or organisation),
- the one sentence in which the term appears, at most 300 characters. We first make other detected personal passages in that sentence unrecognisable (
[PER],[ORG]and so on). If detection is unavailable at that moment or the position in the text is not known, we store no sentence, only term and type, - whether the term was detected wrongly or missed, and the time,
- the identifier of the organisation in which you shared the example. It serves only to delete the examples together with the organisation and to handle a deletion request. We do not take it into the training dataset,
- as proof of your consent (Art. 7(1) GDPR), which version of this policy you consented to.
We do not store your account, session or message identifiers or the mapping of placeholders to terms. We therefore do not link an example to a person.
Purpose: We use the examples to improve our own detection model, including by training it with them. The examples are kept on our server in Germany; for training we take them into a training dataset on our own computers. There we replace the blanked passages with invented substitute values. We do not pass the examples on to anyone and do not use them for any other purpose, including advertising.
Who is responsible: For the shared examples we are the controller ourselves, even if you share them in an organisation. There we take them out of our processing on behalf of the organisation only because the organisation instructed this with the switch (Art. 28(3)(a) GDPR, Data Processing Agreement Section 3a). For the other content of an organisation, Section 2 still applies: there we process it on behalf of the organisation.
Legal basis: your consent, Art. 6(1)(a) GDPR, given by the tick or the report. It is voluntary; without it you can use the Nema services and pseudonymisation unchanged. If other people appear in the sentence (for example a name that detection missed), we process their data on the basis of our legitimate interest, Art. 6(1)(f) GDPR: detection is meant to protect personal data more reliably, including that of these people. Only the one sentence is stored, without reference to the person who shared it and without disclosure. These people can object to the processing under Art. 21 GDPR (support@nema.im).
Withdrawal: You can withdraw your consent at any time with effect for the future (Art. 7(3) GDPR). For future reports it is enough not to tick the box and not to report anything as “should have been detected”. If you want us to delete examples you have already shared, write to support@nema.im. There is no separate way to do this in the app. Because we do not link examples to an account, please give us the organisation, the term and roughly the time (Art. 11(2) GDPR); we will then delete the matching examples. If the organisation switches the switch off, its members can no longer share examples from then on; switching it off does not delete examples already shared. Withdrawal does not affect what was done lawfully until then: we do not reverse a model that has already been trained with an example. From training datasets we still use, we remove the example for future training.
Storage period: We delete the stored example at your request, otherwise together with the organisation. Because we do not link it to an account, it remains even after your account is deleted, as long as the organisation exists. In training datasets it remains as long as we use the dataset for training. For backups see Section 17.6.
What you should know: The example is not anonymised. The term is often itself a name or other personal data, and a sentence can contain clues that our detection does not find; those are exactly the passages it misses. A trained model can reproduce parts of its training examples; this cannot be ruled out entirely. Share only examples whose content you are entitled to dispose of, and no health data or other special categories of personal data (Art. 9 GDPR).
7.2 Field masks
Independently of pseudonymisation, Admins can hide individual fields of events from modules or replace them with an identifier before an agent sees them, such as the sender address or phone number of an email. The identifier is formed with a secret key from the field value and the organisation; it cannot be reversed without this key. Field masks apply to webhooks, regular queries and incoming emails, not to your own MCP servers. No masks are set by default.
Legal basis for both: Art. 6(1)(b) GDPR; Art. 25, 32 GDPR.
7.3 Feedback on detection
If something is replaced as personal data that is not, members of an organisation can report this in Crew: “This is not personal data”. The term is then put on the organisation’s exception list and is no longer replaced in this organisation, in new and in earlier conversations. It then goes to the language model in plain text, even if it turns out to be a name after all; check before you report. We store the term, the type it had been detected as (such as person or place), the time and the member who reported it. Every member of the organisation can view the list and remove terms again. If the organisation adds the term to its own list (Section 7.1), we remove it from the exception list. The list is a setting of the organisation; we process it on its behalf (Section 2) and delete an entry when it is removed or the organisation is deleted.
8. Workshop, previews and SSH keys
Workshop: In the plans that include it, there are isolated environments (containers) with data storage on our server, in which agents edit files and run commands. They hold the files that you or agents create or upload, and the logs of the commands. Each person has a personal workshop and, in each organisation, a separate workshop for that organisation; an agent sees only the workshop of its context. The histories of Claude Code and Codex are also separated by context; the sign-in with Anthropic or OpenAI applies to you as a person. The organisation is responsible for the data in an organisation's workshop (Section 2). If you leave an organisation, your workshop there is paused and deleted after 30 days; it does not pass into your personal workshop, and after leaving you no longer receive its archive; only the organisation does. A container that has not been used for 7 days is removed; the data storage remains.
Previews: Previews are initially visible only to you and the members of the agent's organisation. Via a share link, anyone who knows it can access it without signing in until it expires (1 hour, 24 hours or 7 days) or you revoke it. A server running in the workshop is, for technical reasons, also reachable at an address containing a randomly generated key that cannot be guessed; anyone who knows this address can access it without signing in. What you make accessible via a preview can be read by third parties. When a preview is accessed, we process the data of the request (Section 13.1) and the time of the last access in order to stop unused previews (servers after 2 hours, static files after 30 days).
SSH keys: For each agent we can generate a key pair so that it can access your repositories via Git. We store the private key encrypted in our database. For a run, we place it in the workshop in a directory created only for that run and readable only by the agent; after the run it is overwritten and deleted. In addition, each workshop generates its own key pair on first start, whose private key lies unencrypted in the workshop's data storage until the workshop is deleted.
Legal basis: Art. 6(1)(b) GDPR.
8a. An agent's own screen
For an agent with a workshop you can switch on an own screen (default: off; only the person who owns the agent can switch it). A virtual screen with a Chromium browser then runs in the workshop, which the agent operates with mouse and keyboard.
- Browser profile: Chromium keeps logins, cookies, history, cache and passwords saved in the browser between runs. The profile is stored in the workshop's data storage; passwords saved in it are not additionally encrypted by an operating system. We delete the profile when you delete the agent, at the latest with the workshop (Section 17.4). It is not included in the workshop archive (Section 18) and is not backed up by us. Files the agent downloads in the browser end up in the workshop and are visible to the other agents of the same context.
- Live view: Only the person who owns the agent can watch what is on the screen live in the apps, in an organisation too; its other members cannot. The images go via our server to the watching device; we do not store them. We log when the view is opened; we delete these logs after at most 14 days (Section 13.1).
- Taking over: The person who owns the agent can take over the screen. Mouse movements, clicks, keystrokes and, when pasting, the text of your clipboard (up to 4,000 characters) then go into the agent's screen. We do not log the content of these inputs; we record when the screen was taken over and handed back. The agent is told that, and for how long, you took over.
- Screenshots to the model: After every action the agent takes a screenshot. In the Nema runtime it goes to the model for image analysis (Section 6.1); in Claude Code and Codex it goes as an image to Anthropic or OpenAI (Sections 6.2 to 6.4). Only what Section 7.1 describes is pseudonymised. In the Nema runtime we store the screenshots with the conversation history (Section 17.1).
- Steps: The address the agent opens and where it clicks appear in the steps of the run (Section 5). Of text it types, including passwords, the steps and the history show only the length, not the text.
For logins, captchas or second-factor confirmations the agent is instructed to ask you instead of solving them itself.
Legal basis: Art. 6(1)(b) GDPR; for third-party data on visited websites Art. 6(1)(f) GDPR, limited to the task you give the agent; in organisations on behalf of the organisation (Section 2).
9. Operating your computer
If you switch it on in the Nema Crew app for Mac or Windows and enable it for an agent, the agent can take screenshots, read windows and controls and, at the “Operate” level, also use the mouse, keyboard, files and command line of your computer (Mac or Windows PC).
- Screenshots go to our server and to the language model of the respective runtime. They are pseudonymised only if the agent belongs to an organisation that has switched on pseudonymisation including images (Section 7.1), and even then only what detection finds in the text of the image. Otherwise the model sees everything currently visible on the screen, including messages from other people, notifications, passwords in plain text.
- Output of commands and files read also go to the model.
- Files into the workshop: At the “Operate” level an agent can fetch a file from your computer into its workshop. The app asks you, showing path and size, unless you have set “confirm nothing”. The file goes via our server into the workshop; we delete an intermediate copy in our database as soon as the transfer is completed or aborted. The path on your computer goes to the model.
- Files onto your computer: The app saves a file an agent sends you (Section 4) in your Downloads folder after you have confirmed this. It does so without asking only if you have set “confirm nothing” and the agent has the “Operate” level.
- The app stores a log of actions only on your computer. What a step produced goes as a result to our server and to the model; that a device has saved a file is recorded on the message (Section 4).
While an agent is using your screen, the app shows an indicator with its name and the current step. It is generated on your computer and excluded from screenshots; all that goes to our server for it is whether it should be shown large or small.
Close confidential windows before you let an agent view your screen.
Legal basis: Art. 6(1)(b) GDPR. For data of third parties on your screen: Art. 6(1)(f) GDPR, limited to the task you initiated.
10. Mailboxes for agents
If you store the credentials of one or more email mailboxes (IMAP/SMTP) for an agent, we store them encrypted. This works with any mail provider that offers these protocols; with some, only with an app password that you create there. We do not use a separate programming interface of the mail provider for this. Mailboxes for agents are not a module; Section 11 does not apply to them. The agent retrieves new emails about every minute; we process their content, sender and recipients so that the agent can respond to them. By default, the agent creates replies as a draft for approval; only after your approval do we send them via your mail server. Sent emails always carry a notice that they were created with the help of AI.
The senders of such emails are mostly third parties. For them: your email is read and answered by an AI agent on behalf of the mailbox owner. This information under Art. 13, 14 GDPR is generally owed to the senders by the mailbox owner as controller.
Legal basis: Art. 6(1)(b) GDPR; for the senders' data Art. 6(1)(f) GDPR (handling their request on behalf of the mailbox owner) or data processing on behalf (Section 2).
11. Modules and your own MCP servers
11.1 Principle
Modules connect Nema Crew to other services, such as a calendar, a file store or a ticket system, or give agents additional tools. Modules are optional. A module only processes data once it has been connected or switched on, either by you or, in an organisation, by an Admin or Owner for the organisation. The app shows which modules exist. This policy therefore does not name individual modules.
A connected module exchanges with the service the data needed for the task. It reads that data, and it also writes if you set it up that way. It receives events from the service via webhooks or queries them regularly. Such events can trigger an agent (Section 5).
11.2 Each module's privacy notice
Each module has its own privacy notice. It states:
- who provides and operates the module, with address and contact details, and in which role (Section 11.3);
- what data the module retrieves or sends and what permissions it requests from the connected service;
- what it uses the data for, with its functions and triggers;
- whom the data goes to, including countries outside the EU;
- what the module stores, how that is protected and how long it is kept;
- how you disconnect and how you revoke access at the connected service;
- the legal basis and any specific requirements of the connected service;
- further documents, such as a data processing agreement or the privacy policy of the connected service.
We show you the notice in the app before you connect or switch on a module. We also publish it at crew.nema.im/integrations/‹module name›/privacy. By confirming, you state that you have taken note of the notice. This is not consent (Section 21). A module without a notice cannot be newly connected or switched on.
The notice supplements this policy for the module concerned. It does not repeat what this policy says for all modules.
If a notice changes, a new version is created. We show it to you in the app for confirmation. Existing connections keep running in the meantime. A change to a module's notice is not a change to this policy. We keep earlier versions as proof.
We store which version of a notice you confirmed and when, whether for your account or for an organisation, and with which app. We delete this information with your account. We keep confirmations for an organisation as proof until the organisation is deleted. Legal basis: Art. 6(1)(f) GDPR. Our legitimate interest is being able to show that we provided the information under Art. 5(2) GDPR.
11.3 Who provides and operates a module
-
Modules that we provide and operate run as separate services on our server (Section 20). This policy applies to them, and we are the controller. In an organisation we process the data on its behalf under the data processing agreement (Section 2). The module's notice is our information under Art. 13 GDPR for that module.
-
Modules from other providers are provided by that provider and run on its own systems. The provider receives the data the module needs. Its notice states the provider's name, address and contact details, and those of its data protection officer if it has one. It also states the role in which the provider acts:
- If the provider is an independent controller, its own privacy policy applies to its processing. The notice links to it. We are not responsible for that processing. We only pass the data on to the provider because you or your organisation connected or switched on the module.
- If the provider acts on behalf of your organisation, it does so only on the basis of a data processing agreement between the provider and the organisation. The notice links to that agreement. Such a provider is not our sub-processor.
When you confirm the notice of such a module, you also confirm that your data goes to that provider.
-
The connected service, meaning the account you connect via a module, is not the provider of the module and not our processor. It acts on the basis of your contract with it and under its own responsibility. Its privacy policy applies, and the notice links to it.
11.4 What applies to all modules
- Credentials (OAuth tokens, keys) that you store for a module are kept encrypted (AES-GCM) in our database. This also applies to renewed tokens. The module's notice says whether the module stores any other data itself.
- Disconnecting: You can disconnect any connection in the app at any time. We then delete the credentials from our database immediately, and no agent retrieves data via this module any more. If an Admin or Owner connected a module for an organisation, only the organisation can disconnect it. You revoke an authorisation granted at the connected service there. The module's notice explains how.
- Runtimes: Whatever a module delivers goes as a result or event into the conversation context and thus to the language model of the runtime you chose (Section 6).
- Pseudonymisation and field masks (Section 7) also apply to modules. A module's results and events go to the language model pseudonymised if the organisation has switched on pseudonymisation and has not switched it off for the module. Field masks apply to events from modules. Whatever an agent sends to a module goes in plain text to the connected service or to the module's provider (Section 7.1).
- Content created from module data is stored like other content (Sections 4, 5 and 17). This includes responses, memories, knowledge entries, the steps of a run and cards for events that triggered an agent. It remains stored after disconnecting until you delete it.
- Backups (Section 17.6) and your rights (Section 18) also apply to modules. If a module's provider is an independent controller, direct requests about its processing to that provider.
- In an organisation we process data from modules on its behalf (Section 2). When the organisation connects or switches on a module, that is an instruction to us.
Data from Google APIs: For modules that we operate and that receive data via Google APIs, the following applies: Nema Crew's use and transfer of information received from Google APIs will adhere to the Google API Services User Data Policy, including the Limited Use requirements. The notice of each such module gives the details.
Legal basis: Art. 6(1)(b) GDPR, because the module carries out the connection you set up. For third parties' data that a module delivers, Art. 6(1)(f) GDPR applies, limited to the task you give the agent. In organisations we process on their behalf (Section 2). Each module's notice states its legal basis.
11.5 Your own MCP servers
If you connect your own MCP server, we store its credentials encrypted. What an agent sends to the server goes there in plain text by default. In organisations with pseudonymisation switched on, Admins and Owners can set two things per server in the privacy settings: whether its results go to the model pseudonymised (default: yes), and whether it receives the placeholders instead of the real details when its tools are called (“Send placeholders”, default: no). Your own MCP servers are services that you connect yourself. They are not modules and not our processors. We do not know who operates the server or what it does with the data.
Legal basis: Art. 6(1)(b) GDPR, because we carry out the connection you set up.
12. Notifications, voice and communication
- Email (registration, password, invitations, notices about the contract, verification for the Student plan to your university address, Section 3a): sent via a mail server we operate on our own server.
- Push notifications on iPhone and Android: For a direct message from an agent or when you are mentioned, a notification goes to your device, on the iPhone via the Apple Push Notification Service, on Android via Google's Firebase Cloud Messaging. It contains the name of the agent, where applicable the organisation and the channel and the first 180 characters of the message; for an email draft, recipients and subject; for a request to confirm an action on your computer, device and action; for a file for your device, file name and size. If several accounts use the same device, it states which account the notification is for. Apple or Google receives this text and an identifier of your device for this app (push token) in order to deliver it. On Android, the push component provided by Google also creates an installation identifier and transmits it to Google. For each device we store the push token, platform, app version and language. If you log out or the token becomes invalid, the device receives no more notifications; we delete the entry with your account. You can switch off notifications in your device settings. We delete delivery jobs 14 days after sending.
- Apps for Mac and Windows: The app generates notices itself on your Mac or Windows computer; nothing goes to Apple or Microsoft for this.
- Speech output and speech recognition: If you have a response read aloud or we convert a voice recording into text, this happens on our own server; nothing goes to third parties for this.
- Dictation in web and app via the speech recognition of your browser or operating system: what happens there is determined by its manufacturer.
- Messengers: A connection of the Nema Services to messenger services is not currently active. If we switch such a connection on, we will update this policy beforehand. If you connect a messenger service via a module, Section 11 applies.
Legal basis: Art. 6(1)(b) GDPR; for the push token and installation identifier on your device Section 25(2) no. 2 TDDDG (strictly necessary for the notifications you have asked for and allowed beforehand in the operating system).
13. Websites and apps
13.1 Access
When crew.nema.im, workspace.nema.im, previews and our interfaces are accessed, our server processes, as technically necessary, the IP address, time, requested address, browser and operating system. We keep these server logs for no longer than 14 days.
Legal basis: Art. 6(1)(f) GDPR (secure and stable operation).
13.2 Cookies and local storage
We use technically necessary cookies and entries in local storage: for your sign-in, your language (cookie NEXT_LOCALE, one year) and your settings in the interface (such as light/dark and colour palette). We store the choice of light or dark in the cookie nema_theme (one year); it applies to crew.nema.im and nema.im. To open private previews, we set an access cookie on the preview's address (12 hours; for share links until they expire).
Legal basis: Section 25(2) no. 2 TDDDG; Art. 6(1)(b) GDPR.
13.3 Reach measurement
We measure the use of crew.nema.im with Umami, software that we operate ourselves on our server (analytics.nema.im). Umami does not set cookies. It records the page accessed, referrer, approximate origin, device type, browser and screen size. We do not create a cross-device profile; no data goes to third parties in the process.
We also count downloads of the apps for Mac and Windows and update checks by installed apps in order to see how many installations exist and which versions are in use. For this purpose, we evaluate a log of our download server that contains no IP address, and delete it after the evaluation. So that an installation counts at most once a day, we use the random identifier that the app sends with the update check anyway. Of it, we keep on our server only a check value formed with a daily key, which we delete together with the key at midnight (UTC); the same installation therefore cannot be recognised across days. We pass on to Umami only the type of event (download, update, active installation), the version and the processor type, but no IP address and no identifier. The apps themselves send nothing additional for this.
Legal basis: Art. 6(1)(f) GDPR (measuring reach and the distribution of our apps).
Beyond this, the apps for Mac, Windows, iPhone and Android contain no analytics or crash reporting services. For push notifications the Android app contains components of Google Firebase (Cloud Messaging and the installation identifier it needs), but neither Google Analytics nor Crashlytics.
We do not use third-party tracking services such as Google Analytics, Meta Pixel or advertising networks.
14. Payment
Before your first order, the app asks for your billing details: country, private or business, name or company, address and, where applicable, your VAT or tax identification number. We store them with the payment service provider in your customer account, together with the resulting tax treatment. If you, as a business in another EU state, provide a VAT ID, we check it with the European Commission's confirmation procedure (VIES); the number and our own VAT ID go there, and validity, where applicable name and address, and a request number come back, which we store as evidence.
For paid plans, add-on packages and credit packages we use a payment service provider. There you enter your payment data; card numbers do not reach our servers. We receive a customer ID, the plan and the payment status. In organisations, Admins and Owners book.
Orders, withdrawal and cancellation: For each order we store the order number, content, price and your consents with wording and time (Terms, for consumers early commencement) and send you the confirmation by email. We store withdrawals and cancellations, including those made via the pages crew.nema.im/legal/withdraw and crew.nema.im/legal/cancel without signing in, with their content, the contact details provided and the time of receipt, and confirm them by email. We keep this information as long as claims under the contract are possible and, as business letters or accounting records, for the statutory periods (Section 17.4).
Legal basis: Art. 6(1)(b) GDPR; for invoices, VAT and retention Art. 6(1)(c) GDPR.
14.1 Purchases through the App Store or Google Play
In the app for iPhone or Android you can buy plans, add-on packages and credit packages through Apple's App Store or through Google Play (Terms Section 2.2a). The purchase and payment are then handled by the store: in the EU Apple Distribution International Ltd. (Ireland) or Google Commerce Limited (Ireland). They process your payment data, your store account and the purchase as independent controllers under their own privacy policies: apple.com/legal/privacy, policies.google.com/privacy and the Google Payments Privacy Notice (payments.google.com). They are not our processors. We do not ask for billing details for a store purchase.
What we give the store: With the purchase dialog, our app passes the store a random identifier that we generate for the order and, for Google Play, also the order number (NC-…). We use it to assign the purchase to the account or organisation for which you ordered. We do not give the store your name, your email address or content from Nema Crew.
What we receive from the store: The app passes on to us the store's signed confirmation of the purchase; the store also sends us server notifications when something changes (renewal, auto-renewal switched off, failed payment, expiry, refund, revocation), and we query the status with the store. These contain identifiers of the purchase and the subscription (such as transaction or order ID, purchase token), the product, times of purchase and expiry, the status of the subscription (active, in a grace period, retrying payment, expired), whether it renews automatically, the environment (test or live), our identifier from above, refunds and revocations with time and, where the store includes them, country or region, currency and price. No payment data (card, account) and neither your Apple ID nor the address of your Google account. We store the store's information as we receive it with the order, together with the order number and your consents (above).
Purpose: to activate, renew and end the purchased plan, add-on packages and credits, to implement refunds (for example deducting credits), to prevent duplicate subscriptions through different routes, to restore purchases, to send the order confirmation and to meet evidence and retention obligations.
Retention: like the other order data: as long as claims under the contract are possible and, as business letters or accounting records, for the statutory periods (Section 17.4).
Legal basis: Art. 6(1)(b) GDPR; for retention and tax records Art. 6(1)(c) GDPR.
15. Recipients by category
We pass on data only to service providers we need for the service, and only for the stated purpose. We do not sell data.
We name categories here, not company names, because providers change: a model is discontinued, a better one comes along, a provider changes its terms. Which provider specifically stands behind which category, with its seat, place of processing and basis for any third-country transfer, is stated in the continuously maintained list of providers at crew.nema.im/legal/subprocessors. We announce a change of provider, or a new provider, that processes your content at least 30 days in advance (Terms Section 3.14).
| Category | What goes there | Role |
|---|---|---|
| Data centre operator | all data, on our own server | processor |
| Intermediary service for model interfaces | conversation context, images (Nema runtime, also for selecting the answering agent, Section 5.1) | processor |
| Operators of language models, specified by us per request | conversation context (Nema runtime, also for selecting the answering agent, Section 5.1) | sub-processor of the intermediary service |
| Operator of a model for image analysis, specified by us per request | images, screenshots (Nema runtime) | sub-processor of the intermediary service |
| Payment service provider | payment, billing and tax data | processor; independent controller for its own statutory obligations |
| Apple's App Store, Google Play (only for purchases there) | order identifier, for Google Play the order number (Section 14.1) | independent controller for purchase and payment |
| Apple's push service (iPhone only) | notification text, push token | recipient under Apple's developer terms |
| Google's push service (Android only) | notification text, push token, installation identifier | processor under Google's terms for Firebase |
| Websites an agent opens via its own screen or a module | whatever the agent enters, submits or uploads there | independent controller |
| Search engine (web search of the Nema runtime) | search term | independent controller |
| European Commission's VAT ID confirmation procedure (VIES) | VAT IDs of businesses in other EU states | independent controller |
| Anthropic, OpenAI (only with your own runtime) | conversation context | your contracting party, independent controller |
| Services and MCP servers connected by you | what the task needs | your contracting party, independent controller |
| Providers of modules that we do not operate | whatever the module needs, as stated in its notice | independent controller or processor of your organisation (Section 11.3) |
Authorities receive data only if we are legally obliged to provide it.
16. Transfers to countries outside the EU
Some recipients are established outside the EU and the EEA or process data there, in particular in the USA. For the USA there is an adequacy decision of the EU Commission (EU-US Data Privacy Framework); it covers a transfer only if the recipient is certified under it. Otherwise other safeguards are needed, such as standard contractual clauses of the EU Commission under Art. 46(2)(c) GDPR.
-
Nema runtime: The intermediary service is established and processes in the USA and is not certified under the Data Privacy Framework. We base the transfer to it on the standard contractual clauses that are part of its data processing agreement. It passes the requests on, within the USA, to the operators of the language model for text that we have specified; it is answerable for this onward transfer under the standard contractual clauses. The model for image analysis is operated in the EU.
-
Payment service provider: Our contractual partner is established in Ireland and transfers data to its parent company in the USA. That company is certified under the Data Privacy Framework; standard contractual clauses apply in addition.
-
Google's push service: Google also processes the notifications outside the EU, in particular in the USA. The transfer is governed by Google's terms for Firebase; we will tell you on request to support@nema.im which safeguard applies there.
-
App Store and Google Play: You conclude the purchase yourself with Apple or Google; what they process outside the EU in doing so is governed by their privacy policies (Section 14.1). Only the order identifier goes there from us.
-
Apple's push service: Apple is not certified under the Data Privacy Framework and, by its own account, bases transfers from the EEA on standard contractual clauses.
-
Modules: Whether a connected service or a module's provider processes data outside the EU, and on what basis, is stated in the module's notice (Section 11.2).
You can obtain a copy of the standard contractual clauses we rely on by request at support@nema.im.
Standard contractual clauses bind the recipient, not the authorities of its state. Where the law of the recipient state allows authorities far-reaching access, you have practically no enforceable legal remedy against it. Do not entrust to the Nema Services anything that must not fall into other hands, or switch on pseudonymisation for such terms and observe its limits.
Which provider stands behind which category, with country and basis for the transfer, is stated in the list of providers.
17. Retention period
17.1 Principle: until you delete
We store content (chats, messages, channels, attachments, memories, knowledge entries, agents) without a fixed period, until you delete it, individually or with your account. The reason is the service itself: your agents are supposed to remember what you decided months ago. A deletion period set by us would remove exactly that. Storage is therefore necessary for the performance of the contract under Art. 6(1)(b) GDPR. If you delete a memory (also with “Reset everything” in Memory), it disappears from the app and from the search index immediately; we delete the memory itself and its earlier versions permanently after 30 days at the latest.
In organisations, the organisation decides on the deletion of its content.
17.2 Switch to the Free plan
If the trial period has expired, a plan ends through cancellation or is not paid, the Free plan applies to your personal area (Terms Section 4.7). An organisation without a plan is locked (Terms Section 4.8); until the lock applies to an existing organisation, the Free plan applies to it. Agents, channels, history and memories remain stored; agents beyond the scope of the Free plan, and all agents in a locked organisation, are paused. We keep the data storage of the workshop of the affected context and the histories of Claude Code and Codex there for a further 30 days from the switch to the Free plan or from the start of the lock so that the files can be downloaded, send reminders 7, 3 and 1 day beforehand by email, show the remaining days in the app and delete them afterwards. If a plan is booked again within this period or, for the personal workshop, the Student plan is confirmed, everything is retained. The same applies to your personal workshop if the Student plan expires or is ended and no other plan applies. Workshops in other contexts are not affected.
17.3 Deletion after inactivity
If you do not use an account on the Free plan for 90 days, we delete the account. What counts is the last use with your sign-in, not just signing in itself. For this, we store the time of last use. 14 days beforehand, we write to you by email and point out the export; without this email we do not delete, and any use during this time cancels the deletion. We do not delete accounts that use a booked plan or the trial period, themselves or via an organisation, or accounts to which the Student plan applies, because of inactivity; their agents often work without anyone signing in.
17.4 Individual data
- Account: until the account is deleted
- Credentials of connected services, mailboxes, MCP servers: until you disconnect or delete the account
- SSH key of an agent: until the agent is deleted; the workshop's key until the workshop is deleted
- Earlier versions of an agent's instructions: until the agent is deleted
- Data that a module stores itself: as stated in the module's notice (Section 11.2)
- Confirmations of module notices: until the account is deleted; for an organisation until the organisation is deleted
- Confirmation of an organisation for Claude Code and Codex: until withdrawn or until the organisation is deleted
- Exports: the file 7 days, the information about the export (time, status) 30 days
- Steps of an agent run: 7 days in the cache; up to 50 steps with the agent's response, for as long as it exists
- Delivery jobs for push notifications: 14 days after sending
- Push tokens of your devices: until your account is deleted; logged-out devices receive no more notifications
- Accounts never confirmed: 30 days after registration (Section 3)
- Two-factor sign-in (key, time, time step, recovery codes as hashes): until it is turned off or the account is deleted; recovery codes until new ones are generated (Section 3.1)
- Passkeys (public key and related details): until you remove the passkey or delete the account (Section 3.1)
- Time of “sign out other devices”: until the account is deleted (Section 3.1)
- Failed attempts at the second factor: 16 minutes at the latest; one-time values of the sign-in at most 5 minutes (Section 3.1)
- Verification for the Student plan: as long as it is valid and 12 months after it ends; after a block by us, only the block remains (Section 3a)
- Profile of an agent's own screen: until the agent is deleted, at the latest with the workshop
- Files for your devices and records that a device has saved them: like the message they belong to; signed links expire after about 7 days
- Screenshots in the history (Nema runtime): like the conversation
- Workshop: until you delete the files, at the latest until the account is deleted, 30 days after the respective context switches to the Free plan or after the start of an organisation's lock (Section 17.2) or, for a workshop in an organisation, 30 days after you leave
- Previews: until you delete them; share links until they expire or are revoked
- Consents to Terms: for the duration of the contract and until the limitation period for possible claims has expired
- Detection results (position and type of matches, never text): 60 days without use (Section 7.1)
- Finished pseudonymisation results (text with placeholders): with the conversation or the organisation, otherwise after 45 days without use (Section 7.1)
- Mapping of placeholders per organisation: until the organisation is deleted (Section 7.1)
- Exception list of pseudonymisation: until the entry is removed or the organisation is deleted (Section 7.3)
- Shared detection examples: until you request deletion, otherwise until the organisation is deleted; in training datasets as long as we use them for training; only with your tick or your report (Section 7.1a)
- Backups of the database: up to about 6 months (Section 17.6)
- Invoices and accounting records: 8 years, business letters 6 years, books and financial statements 10 years (Section 147 of the German Fiscal Code, AO; Section 257 of the German Commercial Code, HGB)
17.5 Deleting your account
You can delete your account at any time in the settings. The associated data is then deleted immediately, including your agents, organisations of which you are the only member, attachments, memories, credentials, the entries in the search index, all your workshops with their data storage and the containers for Claude Code and Codex. If the service for the workshop cannot be reached at that moment, we automatically catch up on its deletion. We revoke your agents' SSH keys at the services where they were registered via a connected account. In other people's channels, your messages remain as deleted entries without content or sender. Channels that you created are deleted with all their messages, also in organisations, there including the messages of other members. We end a running subscription through our payment service provider immediately; a subscription through the App Store or Google Play continues with the store until you cancel it there (Terms Section 5.2). We keep order data and invoices for the statutory periods (Section 17.4). Restoration is not possible. The data remains contained in our backups for up to about 6 months (Section 17.6). Export beforehand whatever you want to keep.
17.6 Backups
We back up our database (with accounts, organisations, agents, channels, messages including attachments and screenshots in the history, memories and the data that modules store in our database) completely every day, encrypted (AES-256), on our server and additionally on storage of Hetzner Online GmbH in data centres in the EU. We keep every backup of the last 7 days, then one per week for 4 weeks and one per month for 6 months. Data that you or we delete therefore remains contained in backups for up to about 6 months. We do not remove it there individually; it drops out as the backups expire. Until then, we use the backups only to restore operation after a failure; if we restore a backup, we delete data that had already been deleted again.
Not backed up are the workshops with their files and the browser profile of an agent's own screen, the histories of Claude Code and Codex, the search index and caches. You back up important files yourself via the workshop archive or the export.
Legal basis: Art. 6(1)(b) and (f) GDPR; Art. 32(1)(c) GDPR.
18. Your rights
You have the right to
- access your data (Art. 15 GDPR),
- rectification (Art. 16),
- erasure (Art. 17),
- restriction of processing (Art. 18),
- data portability (Art. 20): in the settings you can download your data yourself as a ZIP file (JSON and attachments) and the data storage of each of your workshops as an archive (without the browser profile of an agent's own screen, Section 8a); whatever is missing there you receive on request,
- object to processing based on Art. 6(1)(f) GDPR on grounds relating to your particular situation (Art. 21),
- withdraw consent with effect for the future under Art. 7(3).
Please contact support@nema.im for this. If your rights concern content of an organisation, we forward the request to the organisation.
You can lodge a complaint with a data protection supervisory authority (Art. 77 GDPR), for example at your place of residence. The authority responsible for us is:
Landesbeauftragte für Datenschutz und Informationsfreiheit Nordrhein-Westfalen
Kavalleriestraße 2-4, 40213 Düsseldorf, Germany
https://www.ldi.nrw.de
18a. Notes for users outside the EU
Wherever you live, we process your data as this policy describes: stored on our server in the EU (Section 20) and passed on only to the recipients in Sections 15 and 16. The rights in Section 18 apply to all users; you exercise them via support@nema.im.
No selling, no advertising. We do not sell personal data, do not pass it on for advertising and do not build advertising profiles. The apps contain no analytics and no advertising services (Section 13.3).
USA, in particular California (where applicable): We do not pass on personal data for money or other valuable consideration ("sale") or for cross-context behavioural advertising ("sharing"); there is therefore no "Do Not Sell or Share" link. You exercise rights that the law of your state gives you, such as access, correction and deletion, via support@nema.im. You will not be discriminated against for exercising them.
19. Security
We protect data, among other things, as follows:
- transmission encrypted via TLS,
- passwords only as bcrypt hashes,
- two-factor sign-in keys encrypted (AES-GCM), recovery codes only as hashes, of passkeys only the public key,
- credentials of modules, mailboxes and MCP servers, SSH keys and jobs in our queue encrypted (AES-GCM),
- each workshop and each own subscription in its own container,
- access to an organisation's content only for its members according to their roles; an agent's own screen can only be watched and taken over by the person who owns the agent.
20. Data centre
Our platform, the database, the workshops, email sending and reach measurement run on a server managed by ourselves from Hetzner Online GmbH (Germany). The server is located in the European Union. A data processing agreement under Art. 28 GDPR is in place with Hetzner. Hetzner has access to the hardware, not to your content during operation.
Legal basis: Art. 6(1)(b) and (f) GDPR.
21. Changes to this policy
We adapt this policy when our services or the legal situation change. We communicate significant changes, in particular new recipients or new purposes, by email. Section 11.2 applies to the modules' notices; they are not part of this policy.
A privacy policy is not a contract. It fulfils our information obligation under Art. 13 and 14 GDPR. You cannot agree to it or object to it, and your continued use is not consent. Where we need consent, we obtain it separately: individually, for a named purpose and revocable at any time.