Cedreon Sub-processors

Version 1.10.1, in force from 2026-10-18

A change takes effect on 2026-11-18. Version 1.11 of this page is published below the text currently in force. It is not yet in effect; the text above governs until that date.
Every third party that processes Customer Content on Cedreon's behalf, what it receives, where it processes, and under which transfer mechanism.

This page is Annex 2 to the Cedreon Data Processing Agreement and forms part of it. It lists the sub-processors Cedreon engages under clause 7 of that DPA, in the sense of Article 28(2) and (4) GDPR and Article 9 of the Swiss Federal Act on Data Protection.

It is written from the deployment as it is actually configured. Where a third party is engaged only when a customer switches an optional feature on, this page says so rather than listing it as though every account used it.

What changed in version 1.10.1. A correction only: no processor was added, replaced or removed, and no row changed. Version 1.10 said that a volume or snapshot of the server decrypts nothing on its own. A snapshot is a copy of the whole server, so it also holds the key that decrypts the stored data, and stored content that is not encrypted, such as conversation titles, which are stored masked. The note in section 1.5 now says so.

Cedreon is operated by SFF S26 AG, St. Gallen, Switzerland.

How to read this page

A sub-processor is a third party that processes Customer Content, meaning the messages, documents and files you submit, on Cedreon's behalf and on your instructions. Those are in section 1, and they are what Annex 2 to the DPA consists of. Section 1.3 also names one further recipient, Anthropic on one model row, which does not serve requests: it is listed there so that this Annex discloses it, and its entry says what it is and is not bound by.

A processor for data where Cedreon is its own controller, such as your billing contact or your login address, is not a sub-processor under the DPA. Clause 2 of the DPA puts that processing in the Privacy Policy instead. Section 2 lists those anyway, because you are entitled to know who touches your data and a list that hides behind a definition is not a disclosure.

Engaging a third party for ancillary services that involve no access to Customer Content, such as telecommunications, connectivity or office facilities, is not the engagement of a sub-processor. See DPA clause 7.

1. Sub-processors of Customer Content

1.1 Model routing

Eden AI (France)

What it does: the single gateway through which every model request is made. Cedreon calls its European endpoint, which routes only to providers cleared for processing in Europe and rejects a request that no such provider can serve, rather than forwarding it elsewhere. Cedreon additionally pins each request to the one Model Provider named below for the model chosen, with provider fallback disabled.
What it receives: the screened prompt. Recognised sensitive values have already been replaced with placeholders before the request leaves Cedreon, so what the gateway forwards carries placeholders in their place. It also receives the model identifier and token counts.
Processing location: European Union.
Transfer mechanism: none required for the EU. Transfers from Switzerland to the EEA rest on the Federal Council's adequacy recognition of the EEA.
Retention: the gateway's published provider data policy states that it stores no request or response content by default, and it logs operational metadata without content. That is a statement about the gateway. What each Model Provider retains is a separate question, answered row by row in section 1.2.
Use: the gateway's data processing agreement states that it does not train on customer data, and binds it for its own processing. The same published provider data policy states that none of the Model Providers named in section 1.2 trains on customer data. That page is the gateway's summary of each provider's policy for the gateway's own accounts. The gateway's agreement does not make it responsible for the Model Providers' practices, and no Model Provider has given Cedreon an undertaking directly; section 4 states what Cedreon remains responsible for.

1.2 Model Providers

These are engaged through Eden AI. Which one serves a given request depends on the model chosen for it, and on any Compliance Profile in force for the account. Every one of them is established in Europe and processes there, and each request is pinned to the single provider for its model with fallback disabled: the request is served by the company named here, or it fails. On one row, Claude Fable 5.1, the provider that serves the request also shares it with the model's maker; that recipient is described in section 1.3, because it does not serve the request and is not a Model Provider.

The maker of a model and the processor that serves it are different companies, and this page names the second. For an open-weight model the two are routinely unrelated; naming the maker as the processor would name a company that never receives the request.

Mistral rows: served by Mistral AI SAS (France), in the European Union. Maker: Mistral AI.
Claude rows: served by Amazon Web Services EMEA SARL (Luxembourg) via Amazon Bedrock, in eu-west-1, Ireland, except the Claude Fable 5.1 row, which is served by Google Cloud EMEA Limited (Ireland) via Vertex AI, in the European Union, and shared with Anthropic as described in section 1.3. Maker: Anthropic.
GPT rows: served by Microsoft Ireland Operations Limited via the Azure OpenAI Service, in the EU Data Zone. Maker: OpenAI.
Gemini rows: served by Google Cloud EMEA Limited (Ireland), in the European Union. Maker: Google.
Qwen 3.8 Flash Next: served by TensorX Ltd (Ireland), in the European Union. Maker: Alibaba Group.
Kimi rows: served by TensorX Ltd (Ireland), in the European Union. Maker: Moonshot AI.
DeepSeek rows: served by TensorX Ltd (Ireland), in the European Union. Maker: DeepSeek.

A correction to versions 1.2 and 1.3 for the Claude Fable rows. Those versions named Amazon Web Services EMEA SARL for every Claude row. Neither of the two Claude Fable rows then offered, Claude Fable 5 and Claude Fable 5.1, was served by it: both were served by Google Cloud EMEA Limited via Vertex AI, which also serves the Gemini rows and was already named on this page, and Claude Fable 5.1, the one still offered, still is. Version 1.4 made this correction, and it is carried forward here.

Retention, row by row. The processors differ, so a single sentence about all of them would be false in one direction or the other. Each position below is stated with its basis.

Zero-retention rows. The gateway's published provider data policy (https://www.edenai.co/docs/v3/data-governance/provider-data-policies, last updated 27 August 2026) lists Amazon Bedrock, Google Vertex AI, Microsoft Azure, Mistral and TensorX as storing no prompt or completion, and advises pinning a request to one of them for zero data retention. Cedreon pins every request to its named processor with fallback disabled. On that basis the Mistral, Claude (other than Claude Fable 5.1), GPT, Gemini, Qwen 3.8, Kimi and DeepSeek rows are zero-retention, with these limits stated rather than implied: operational metadata without content, such as timestamps, token counts, the model and the cost, is kept by the gateway and by the processor; the policy page is the gateway's summary for its own accounts, Cedreon has not verified how each processor account is configured, and a processor's own terms may allow it to log content its abuse classifiers flag; and none of it is an undertaking any processor has given Cedreon.
Exceptions on zero-retention rows. On the Amazon Web Services rows, if Amazon Bedrock detects apparent child sexual abuse material in an image input, it blocks the request, may store and review the flagged input or output solely to determine whether it is such material, and may report it to the US National Center for Missing and Exploited Children or a relevant authority; any such report is made by Amazon, not by Cedreon. On the Gemini rows, Google holds an in-memory cache for up to 24 hours unless disabled, which Google states is compatible with zero data retention. TensorX keeps usage metadata without content for up to twelve months.
Claude Fable 5.1: not zero-retention. Google logs every prompt and completion on this row for up to 30 days for abuse monitoring, and shares that data with Anthropic, which keeps it for at least 30 days. See section 1.3.

TensorX Ltd (Ireland), company number 796387, registered office Unit 25, Classon House, Dundrum Business Park, Dublin 14

What it does: serves the Kimi K3, Kimi K2.7 Code, Qwen 3.8 Flash Next, DeepSeek V4 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro rows, engaged through Eden AI in the same way as every other Model Provider above.
What it receives: exactly what the other Model Providers receive, described in section 1.1: the screened prompt with recognised sensitive values already replaced by placeholders, the model identifier, and token counts.
Processing location: Dublin, Ireland and Helsinki, Finland. These are the only two facilities named in TensorX's own contractual documents. Marketing material elsewhere mentions a third location; we state the contractual position, not the marketing one.
Transfer mechanism: none required. The entity is established in Ireland and both facilities are in the EU.
Retention: TensorX's data processing agreement states that it does not store, log or retain prompts or completions, and its privacy policy that it keeps usage metadata without content for up to twelve months. The gateway's published provider data policy lists it as storing no prompt or completion.
Use: TensorX's own data processing agreement and privacy policy each state that it does not train on customer content. This is in addition to what the gateway states in section 1.1. There the serving processor is TensorX, not the model maker: a note on the gateway's policy page that DeepSeek and Moonshot AI may train on data concerns those companies as providers, and these rows do not reach them.
Why it replaced Nebius B.V. for the Kimi rows: on 11 September 2026 the gateway still listed the Kimi models as cleared for European processing through Nebius B.V., and served them. By 13 September it had reclassified every Nebius model as outside its European set, and requests to them are now refused. TensorX is the only provider on the European endpoint that carries these models. The alternative was to withdraw them, and this page would say so if that were the choice we had made.
A correction to version 1.2 for the Qwen 3.8 rows. Version 1.2 recorded those two rows as served by Alibaba Cloud. Alibaba Cloud's European route does not carry that model generation, so no request to those rows could be served and none was: the gateway refused them, and they returned an error rather than an answer. No Customer Content was processed by anyone under that entry. It was a description of a route that could not run, and it is corrected here rather than left to stand.
What we state plainly about it: TensorX is a recent company, incorporated in September 2025, and it holds no completed ISO 27001 or SOC 2 certification at the date of this page. We say so because the other providers named above are long-established and a reader is entitled to know that this one is not, and because the alternative is to let a uniform list imply a uniformity that is not there. What section 1.1 says applies to it identically: it processes in Europe, and the gateway states that it does not train on customer data.
Withheld from organisations under the EU-only Compliance Profile, which reaches only the Mistral rows, and flagged as provisional under both professional-secrecy profiles: its retention and training position is stated above, while its confidentiality and sub-processor terms, and the contractual chain from Eden AI to it, are not yet verified.

Transfer mechanism: none required for the processors that serve a request. Every one of them is established in the EU or EEA and serves the request there. Transfers from Switzerland rest on the Federal Council's adequacy recognition of the EEA. The sharing with Anthropic on the Claude Fable 5.1 row is the exception, and is described in section 1.3.
A European processing region is not a European maker. Most of the models above were built outside Europe. An organisation that requires the model to have been built by a European company as well as run by one can use the EU-only Compliance Profile, which restricts the account to the Mistral rows.

The current mapping of model rows to processors is the list above. The model menu inside the product groups models by maker and states that every model runs on European infrastructure; an organisation under a Compliance Profile additionally sees the approved route set, with the processor, the training position and the retention position for each row, on its agreements page in the console.

1.3 A further recipient on one model row

Anthropic, the maker of the Claude models, on the Claude Fable 5.1 row only

Why it is on this page: Anthropic requires 30-day retention of all traffic on this model, on every platform that offers it, and Google enables the model only with data sharing to Anthropic switched on. So on this row, and on no other, the request reaches a company beyond the processor that serves it. Anthropic does not serve the request.
What it receives: the prompts and completions on that row as Google holds them: the screened prompt, with recognised sensitive values already replaced by placeholders, and the reply the model returned, which carries the same placeholders. It receives nothing from any other row.
Use: Google states that this data is logged for the sole purpose of monitoring for abuse and is not used to train or fine-tune any model. Anthropic states that it will not use this data to train new Claude models or for any purpose not related to safety, that by default none of its personnel can read it, and that a small set of approved reviewers reads content only where Anthropic's automated trust and safety systems flag it, with every access logged.
Retention: 30 days, and longer only where content has been flagged by those systems or where Anthropic is legally required to keep it.
Processing location: Anthropic states that on Google Cloud the retained data stays in Google Cloud. Neither company states the region in terms Cedreon has seen, and Cedreon has not verified it. Nor do they state which Anthropic entity accesses the data, or from where its reviewers do; Anthropic, PBC is established in the United States.
Transfer mechanism: Cedreon holds no contract with Anthropic. The sharing is a condition Anthropic sets and Google applies under Google's own terms for these models, and Cedreon cannot state which transfer mechanism covers access from outside the EEA. We say so rather than name one we have not seen.
Obligations: Cedreon has imposed no data protection obligations on Anthropic, by contract or otherwise. Its handling of this data rests on the terms between Google and Anthropic and on the published statements above. The undertaking in section 4 covers the processors Cedreon engages and does not reach Anthropic.
How it is avoided: the automatic model choice never selects this row. It is reached only when somebody names it, or names Claude Fable 5, the row it replaced, which is answered on it: by choosing it in the model menu, on a plan that includes it, or by naming it in a request made with an API key, which is not limited by plan. Naming any other model means Anthropic receives nothing.

1.4 Engaged only when you switch a feature on

Both of the following are off by default, are latched per conversation, and are re-resolved on the server, so a modified client cannot enable them for itself.

Linkup (France)

What it does: web search, used only in a conversation where web search has been explicitly switched on.
What it receives: the search query as you typed it. This is deliberate and it is disclosed on the toggle itself. Masking the query does not degrade a search, it changes it: a question about a named company, sent with the name replaced, returns results about something else, and the model is then handed those results as though they answered you. Switching web search on is therefore the disclosure, and the feature is off until you do.
What it does not receive: your documents, your conversation history, and any part of a message other than the question being searched.
Processing location: European routing, not yet contractually pinned. Linkup is a French company and Cedreon's search requests are routed to its European serving. A contractual commitment pinning the serving region to Europe is being put in place; until it is signed, this entry states the operational position and claims no pin. A vendor's country of establishment is not the location of processing, and this page does not treat it as one.
Transfer mechanism: Standard Contractual Clauses (Implementing Decision (EU) 2021/914) with the Swiss addendum recognised by the FDPIC, for any processing outside the EEA.
Withheld entirely from organisations under the EU-only Compliance Profile, and under both professional-secrecy profiles, because the processing region is not pinned.

Google LLC (United States), for connected Drive and Gmail sources

What it does: read-only search of a Google account that the account holder has explicitly connected through Google's own consent screen.
What it receives: a search query derived from your question, destructively masked before it is sent. Google is already the custodian of the files being searched.
Processing location: Google's infrastructure.
Transfer mechanism: Google's data processing terms incorporating Standard Contractual Clauses.
Withheld from organisations under the EU-only Compliance Profile for the in-chat search path.

1.5 Infrastructure

Amazon Web Services EMEA SARL (Luxembourg), operating AWS Lightsail

What it does: hosts the application and its database.
What it receives: everything the service stores, in the form in which it is stored.
Processing location: eu-central-1, Frankfurt, Germany. Daily automatic instance snapshots are taken and held in the same region.
Transfer mechanism: none required. The contracting entity is European and the region is European.
Note on what a snapshot contains: a snapshot is a copy of the whole server. Stored message copies, the placeholder mapping and stored integration credentials are encrypted by Cedreon before they reach the disk, but the key that decrypts them is held on the same server, so a snapshot holds the key as well, together with stored content that is not encrypted, such as conversation titles, which are stored masked.

Cloudflare, Inc. (United States), contracting through Cloudflare Germany GmbH for EEA customers

What it does: authoritative DNS, TLS termination at the edge, DDoS protection, and the outbound-only tunnel that carries requests to the application. Nothing listens publicly on the server itself.
What it receives: inbound requests transit Cloudflare's edge, where TLS is terminated. Request content is therefore processed in the clear at the edge, transiently and in memory, before being re-encrypted to the origin. We state this plainly because it is true of every service behind a reverse proxy and is usually left unsaid.
Processing location: Cloudflare's global network. Requests from Europe are ordinarily served by European points of presence.
Transfer mechanism: Cloudflare's data processing addendum incorporating Standard Contractual Clauses.

2. Processors for data where Cedreon is the controller

These are not sub-processors under the DPA. They do not receive Customer Content. They are listed for completeness.

Stripe Payments Europe, Ltd. (Ireland)

What it does: subscription billing, checkout and the customer billing portal.
What it receives: the billing contact, the subscription record, and payment details entered directly into Stripe's own hosted forms. Card details never reach Cedreon's servers.
Processing location: European Union, with onward transfer to Stripe, Inc. under Standard Contractual Clauses.

Transactional email relay

What it does: delivers account email, meaning address-verification links, seat invitations and replies to contact enquiries.
What it receives: the recipient address and the message Cedreon sends. It never receives Customer Content.

Umami, self-hosted

What it does: website analytics, on the public marketing pages only.
Where it runs: on Cedreon's own infrastructure, in the same Frankfurt region, in a container we operate. No third party receives it.
Where it does not run: it is never loaded inside the signed-in workspace. The console makes no third-party requests at all.

3. Not sub-processors

These run inside Cedreon and send nothing outward. They are listed because customers ask, and because "no third party is involved" is a stronger statement than silence.

Sensitive-value recognition. The recognition model runs locally, on our own infrastructure, as part of the application process. No content is sent anywhere to be screened.
Document text extraction. Word, PDF and spreadsheet text is extracted in-process.
Optical character recognition for scanned documents runs locally as a subprocess at ingest time.
Search indexing and embeddings for the document library are computed locally, over the already-pseudonymised copy.
PDF generation is rendered in-process and offline.
GitHub, Inc. holds the source code repository and runs the deployment pipeline. It holds no Customer Content and has no access to the production database.

4. Changes to this list

A change is published here at least 30 days before it takes effect, and the account contact of each organisation is notified.

If you reasonably object on data-protection grounds within that period, we will work with you in good faith to find a solution. If none is found, you may terminate the affected part of the service without penalty and receive a pro-rata refund for the unused remainder of the paid period. This mirrors DPA clause 7, which is the operative text.

Where we engage a sub-processor, we impose on it data protection obligations equivalent to those in the DPA, by contract or another legally binding instrument. Cedreon remains fully liable to you for a sub-processor's performance of those obligations. The exception is Anthropic on the Claude Fable 5.1 row (section 1.3), on which Cedreon has imposed no obligations; that entry states what its handling rests on.

Questions about anything on this page: [email protected].

Coming into force on 2026-11-18

Version 1.11, published and not yet in force: it takes effect on 2026-11-18.

What changed in version 1.11. Cedreon's server moves to Switzerland, and Cloudflare leaves the path of requests to the service. From 18 November 2026 Infomaniak Network SA (Switzerland) hosts the application, its database and their backups, in place of Amazon Web Services EMEA SARL on AWS Lightsail in Frankfurt, Germany (section 1.5). The encrypted connection now ends on Cedreon's own server, so no third party in front of it processes request content. Cloudflare, Inc. is therefore no longer a sub-processor of Customer Content: it keeps only the DNS for cedreon.com, and is listed in section 2. Amazon Web Services EMEA SARL remains a sub-processor for the Claude rows it serves (section 1.2). Nothing else on this page changed. This version is published at least 30 days before it takes effect, as section 4 requires.