Questions and answers
Frequently asked questions about Kisuno
41 questions on data protection, operation, features, permissions and licensing, each with a complete answer. Kisuno is ready for production use; access to the managed Swiss instance is allocated in stages because demand is high.
In brief
- Kisuno is operated in ISO 27001 certified data centres run by Green in Lupfig (Switzerland); the certificates are held by Green, and an ISO 27001 certification of Kisuno itself is in progress.
- Customer data does not train any models, telemetry is disabled by default, and every tool call through the MCP gateway is recorded in an exportable audit log.
- The models from OpenAI (ChatGPT) and Anthropic (Claude) run in the AWS data centre in Frankfurt; for higher data-protection requirements, open frontier models such as GLM, Kimi and DeepSeek are available from European providers on servers in Europe.
- Kisuno will be released under the MIT licence shortly: self-hosting is free from a single user, with no seat limit and no forced branding.
- Permissions reach down to individual tools and are evaluated in the order user, role, group, server default.
Data and privacy: where your data is held and what happens to it
Kisuno is operated in ISO 27001 certified data centres run by Green in Lupfig (Switzerland), the models run in the AWS data centre in Frankfurt or with European providers on servers in Europe, and customer data does not train any models. Anyone who hosts the software themselves keeps all data on their own infrastructure.
Where is my data processed?
Conversations, files and accounts are held in the data centres of Green in Lupfig, Switzerland. The models from OpenAI (ChatGPT) and Anthropic (Claude) run in the AWS data centre in Frankfurt; for customers with higher data-protection requirements, open frontier models such as GLM, Kimi and DeepSeek are available, operated by European companies on servers in Europe. Anyone who self-hosts keeps all data entirely on their own infrastructure. The data centres are certified to ISO 27001; the certificates are held by Green and not by Kisuno, and an ISO 27001 certification of Kisuno itself is in progress. The supporting evidence is available in the Trust Center.
Is my data used for training?
No. Whatever model you choose: inputs and uploaded data are neither stored by the model provider nor used for training; the data processing agreement sets this out in the section "No training on customer data". One exception we state openly: AWS abuse detection can flag messages, for instance on signs of attacks, malicious code or prohibited content. AWS retains flagged messages with the GPT models and all requests with Claude Fable 5 and 5.1 for up to 30 days, to review and stop misuse; none of it is shared with the model providers or used for training, and the source for this is given in the Trust Center. Telemetry is switched off by default in all components, and no tracking takes place without active consent.
Is our data transferred to the USA?
They are not transferred: conversations, files and accounts are held in Switzerland, and the models from OpenAI (ChatGPT) and Anthropic (Claude) run in the AWS data centre in Frankfurt. To be open about it: AWS is a US company, so the US CLOUD Act applies even with servers in Frankfurt. Anyone who wants to rule that out chooses the open frontier models such as GLM, Kimi and DeepSeek, which European companies operate on servers in Europe; the CLOUD Act does not apply there. We name two further exceptions openly: web search forwards search queries to external search engines without any user identity, and the administration can additionally connect a deliberately flagged provider outside Europe. Anyone who self-hosts chooses every provider in the chain themselves.
Does all data really stay in Europe?
For storage and inference yes, for search queries no. Conversations, files and accounts are held in Switzerland, and the models run in the AWS data centre in Frankfurt or with European providers on servers in Europe. Web search, by contrast, passes the search query on to external search engines, without the identity of the user. The legal space remains a question of its own: AWS is a US company, so the CLOUD Act applies in Frankfurt too. That is why we do not claim across the board that everything stays in Europe.
What happens to our chats if we cancel?
Conversations are held in the database of your own instance and can be exported before shutdown, after which they are deleted. With self-hosting you keep the database anyway and do not have to ask anyone to hand it over. Deadlines and procedures for return and deletion are governed by the data processing agreement.
How long are conversations stored?
Out of the box no retention period is set; one can be configured freely in the admin panel, after which expired content is removed together with its search index entries. Without a set period, conversations remain stored until users or the administration delete them. Exporting and deleting the data of an individual person is likewise carried out by the administration in the admin panel, including full deletion for erasure requests.
Can employees accidentally give company data to the AI?
Yes, employees can type confidential text into any chat tool, including Kisuno. The difference lies in what happens next: the input stays in the instance of your company, does not train any models and is logged. An approved company instance therefore usually works better than a ban that pushes employees onto private accounts.
Do we need a data protection impact assessment?
That depends on the use case, not on the tool: under the GDPR and the revised Swiss FADP an impact assessment is required when processing is likely to result in a high risk to the rights of the data subjects. For that review, Kisuno provides the data processing agreement, the subprocessor list and details of the hosting; a document on the technical and organisational measures is in progress. This answer is not a substitute for legal advice.
Who has access to our data?
Within your instance, users see exactly what their role and group allow, and every access to a connected system is logged. Anyone who self-hosts grants no third party access. In managed operation the operator accesses the system only for maintenance, as governed by the data processing agreement and the subprocessor list.
Does Kisuno collect telemetry or tracking data?
No, telemetry is disabled by default in all components, and no tracking takes place without active consent. Only what you need yourself is logged: logins, administration actions and tool calls in the audit log, plus consumption for the cost overview. This data stays in your instance.
Operation and hosting: who runs the platform and what it requires
Kisuno runs managed in Switzerland, in your own cloud or on your own hardware, always with the same software and without a separate hosting build. Self-hosting follows shortly with the release of the source code; the hosted offering is released, and access to it is allocated in stages because demand is high.
Which operating models are available?
Three: managed by us in Switzerland, in your own cloud or entirely on your own hardware. In all three cases the same software runs in the same architecture; there is no stripped-down self-host edition. The hosted offering is available; because demand is high, we allocate access in stages via a request. Testing is free and not limited in time: every new account starts with a starting credit of CHF 50 (EUR 55) for model usage, and no payment details are required.
Does this work without an IT department of our own?
Yes, if you choose the managed option: operation, updates and backups are then handled by the operator, and you only manage users, roles and permissions in the admin panel. For self-hosting you need someone who can set up a Linux server with Docker and keep it up to date. An IT service provider can also take over operation for you.
How quickly is Kisuno set up?
The technical part takes minutes: obtain the package, run an installation script, open the browser. What comes afterwards takes longer: creating users, defining roles and permissions, connecting systems and working through the first use case properly. Plan on days rather than minutes for a productive rollout. The technical procedure is described on the page about self-hosting.
What server do we need for self-hosting?
A small Linux server is enough, because the computing load sits with the model provider and not with you. Measured on the running stack, the base load is around 0.7 GB of memory and the CPU load close to zero in normal operation; budget 1.5 to 2 GB per instance including code execution. For small teams a machine with 4 vCPUs, 16 GB of RAM and an 80 GB disk is enough to run several instances.
Does it also work without internet?
The platform itself runs entirely inside your network, the language models generally do not: requests to a model provider need an internet connection, and web search does in any case. Fully offline operation is only possible with locally hosted models, which requires suitable graphics cards. Such models can be registered as a provider; the model size that runs on your hardware then determines the result you can achieve.
Can several companies run on one server?
Yes, but each company gets its own, fully separated instance: its own database, its own files, its own credentials, its own network and its own environment for code execution. Only uncritical elements such as the container images and an upstream reverse proxy are shared. Up to around 15 users several companies share one machine; above that each company gets its own, as is also the case in regulated industries, where dedicated infrastructure is contractually assured, or when an in-house ERP system is connected.
How do updates and backups work?
Each instance is updated individually, and to a firmly pinned, previously tested version rather than a moving target. A supplied script backs up the database, the configuration and uploaded files into an archive. The same script restores an instance on a different server, which makes a migration a routine operation.
Can we switch from managed operation to self-hosting later on?
Yes, and the other way round as well, because managed operation and self-hosting use the same software in the same architecture. The switch runs via backup and restore on the target server. There is no licensing hurdle to slow a switch down: Kisuno will be released under the MIT licence shortly.
Features and models: what Kisuno delivers day to day
Kisuno brings together the current models from OpenAI (ChatGPT) and Anthropic (Claude) as well as open frontier models, web search with source references, file generation, a memory and the connection of your own systems in a single interface. The answers below say what of that is running today and what is planned.
What is the difference from ChatGPT in the browser?
ChatGPT in the browser is a personal account with no connection to your company systems and no central control; Kisuno is a company instance with roles, permissions, an audit log and a cost overview. The models are partly the same; the difference lies in the data path and in the controls. On top of that comes the connection of your own systems, so that answers are based on company data rather than on pasted excerpts.
Which models are available?
Kisuno offers the current models from OpenAI (ChatGPT) and Anthropic (Claude); they run in the AWS data centre in Frankfurt. AWS is a US company, so the CLOUD Act applies even with servers in Frankfurt, and we say so openly. For customers with higher data-protection requirements, open frontier models such as GLM, Kimi and DeepSeek are available, operated by European companies on servers in Europe; the CLOUD Act does not apply there. The administration can curate the selection and set a default model, so that users are not faced with a long list.
Can we use our own models or our own API keys?
Yes, Kisuno is model-agnostic: any OpenAI-compatible endpoint can be registered by configuration, including your own keys. Locally hosted models are connected the same way. There is no technical lock-in to a single model provider.
Can I connect my ERP system?
Yes, via the MCP gateway. Permissions apply per individual tool and not just per system, and every call is logged. If the identity of the person is also passed through to the ERP system, the existing user permissions there continue to apply; the gateway does not override them. Which systems are prepared is listed on the integrations page.
What is an MCP server?
An MCP server is a bridge that lets an AI platform operate another system; MCP stands for Model Context Protocol, an open standard. The server provides individual tools, for example "read invoice" or "search customer", and permissions are granted at exactly that level. Ready-made MCP servers exist for HubSpot, Slack, Jira, Confluence, Notion, Linear, GitHub, Sentry, PostgreSQL, Stripe and Google Drive; for your own systems a REST API is enough.
Can Kisuno generate Excel, Word and PDF files?
Yes, as soon as the code sandbox is activated for your instance; that is part of the setup. An instruction such as "generate an Excel file for me" is enough, and the finished file is then available for download. The code required for this runs in an isolated environment, not in the browser and not on the machine of the user. An egress proxy with an allow list and additionally hardened isolation of that environment are in progress.
How does Kisuno obtain up-to-date information from the internet?
Through a built-in web search that backs answers with source references and runs on a self-hosted SearXNG. To be open about it: the search query itself is passed on to external search engines, though without the identity of the user. Anyone who does not want that switches web search off.
Does Kisuno remember what matters for our work?
Yes, Kisuno keeps a memory and independently records what will be useful for later answers. The stored entries can be viewed, deleted individually and reside on your own server. Anyone who wants nothing stored at all can disable the feature.
Is there an app?
Today Kisuno can be used as an installable PWA on iPhone, Android devices and the desktop; it launches with its own icon like an app. Native apps for iOS and Android are planned, though no date has been set.
Is Kisuno available directly in Outlook?
Yes, as an add-in in beta. It opens as a pane next to the open email and can summarise, draft a reply, translate between German and English and extract tasks with owners and deadlines. Today it is in internal use at our company and not yet generally available; the rollout will follow together with the Excel add-in. Until then everything runs through the browser, with the same range of functions, only the route is longer.
Who is liable if the AI says something wrong?
Language models can produce false output, which is why answers are drafts and not verified information; responsibility for using a result remains with the company working with it. Kisuno reduces the risk through source references in web search, through answers based on real company data rather than on the memory of the model, and through a log of every data access. What applies contractually is set out in the terms and conditions; this answer is not a substitute for legal advice.
Permissions, roles and costs: who may do what and what it consumes
Kisuno controls access and costs upstream of the model: permissions apply down to individual tools, every call is recorded in the audit log, and a shared monthly budget caps consumption across the entire organisation.
Who in the company may see which company data?
That is decided by a rule chain upstream of the model, not by the model itself: the check runs in the order individual user, role, group and server default, and the first match wins. If the check fails, access is denied rather than granted. You maintain roles and groups in the admin panel or take them from Entra ID. How this chain is structured in detail is described on the governance page.
How do we prevent the AI from seeing payroll data?
By not releasing the tool in question: permissions apply per individual tool and not per system, so "read invoices yes, salaries no". In addition, the permissions of the source system continue to apply, so that two independent checks take effect. Every call is logged with person, tool and decision.
Where do roles and groups come from?
Maintained locally in the admin panel or taken from Entra ID, depending on whether an identity service is in place. Sign-in via Entra ID or another OIDC provider is supported by the platform and set up per instance; users are then created automatically on first sign-in and their groups are taken over. Automatic provisioning and deactivation via SCIM is planned; today the rule is: no account in the directory, no login.
What does the audit log contain?
Every tool call through the MCP gateway, every administration action and every login, each with person, timestamp and decision. The log can be exported and is therefore usable as evidence for auditors or data protection authorities. Entries are appended continuously and not overwritten.
How do we keep AI costs under control?
Consumption is reported per person, group and model, in tokens and in euros, and a shared monthly budget for the entire organisation caps it. Instead of tight individual limits, all users share the same pot, which avoids frustration about arbitrary boundaries. An additional monthly limit per person can be set if you want one.
What happens when the monthly budget is used up?
A soft block takes effect: requests are stopped instead of quietly continuing, and the administration can raise the budget or release individual people specifically. Exceeding the budget without notice is not provided for. That way there is no invoice at the end of the month that nobody authorised.
What are role agents, and who releases them?
Role agents are prepared AI coworkers for recurring tasks, for example for proposal texts or evaluations, with defined instructions and tools. They are released per group by the administration and not by each user individually. This keeps control over who uses which agent and therefore which data sources.
Open source and licensing: what MIT means in practice
Kisuno will be released under the MIT licence shortly and is a thin fork of LibreChat. Self-hosting is free from a single user, with no seat limit and no forced branding.
Is this really open source?
Yes. Kisuno will be released under the MIT licence shortly and is a thin fork of LibreChat. Self-hosting is free from a single user, with no seat limit and no forced branding. The LibreChat base is publicly available today, is MIT licensed and has more than 40,000 stars on GitHub; the GitHub repository with our extensions and its publication under MIT are coming soon. After that the code can be inspected, modified and passed on, including for commercial purposes.
What does the MIT licence mean in practice?
You may use, modify and pass on the software and build it into your own products, as long as the licence text is retained. There is no user limit, no forced branding and no features locked behind an enterprise barrier. This is where MIT differs from licences that require your own branding or a commercial licence above a certain number of users.
Is self-hosting really free?
Yes, the software is free and has no user limit, starting from a single person. What you bear is your server and the model provider you choose, both independent of us. The managed instance costs CHF 7 (EUR 7.50) per person per month. Through the end of the current month: 50 percent off with a twelve-month term, that is CHF 3.50 (EUR 3.75) per person per month. All details on this are on the pricing page.
What is the difference between Kisuno and LibreChat?
Kisuno is a thin fork of LibreChat, extended by what companies additionally need: an MCP gateway with permissions down to tool level, an admin panel with a cost overview and a shared budget, and a ready-made operating package. The fork deliberately stays close to the original so that new developments from LibreChat continue to flow in. On top of that come German-language support and the option of handing over operation entirely.
Who is behind Kisuno?
Behind Kisuno stands Herr Informatik GmbH in Windisch, Switzerland, which uses the platform internally itself. Anyone who only wants to run the software themselves does not need a contract with us. Why we are building the product is explained on the vision page.
Still have a question?
The hosted Swiss instance is running. Because demand is high, we allocate access in stages; afterwards you can clarify open questions directly on your own instance.
Last reviewed: