# AIMan > Every leading AI model in one workspace. Governed from the first prompt. AIMan gives everyone in a > company one sign-in to Anthropic Claude, OpenAI GPT, Google Gemini and other hosted models, and to > open models that run on dedicated GPU servers in Germany — with answers grounded in the company's > own documents and in current web results, with sources. IT decides which models are on offer, who > may use what, and holds the vendor keys centrally. A product of brandcompete GmbH, Germany. AIMan is an AI workspace and multi-vendor model gateway for companies. People chat with two kinds of model in one interface: open models on GPU servers that brandcompete (or, on-premise, the customer) runs, and the vendors' hosted models, used under the company's own keys. Answers can draw on uploaded documents, web pages and Atlassian systems, or on the web, and cite what they used. Each customer has an installation of its own. - **Contact for sales, pricing and pilots:** sales@brandcompete.com, or the contact form at the bottom of the home page (https://www.aiman.rocks/#contact) - **Vendor:** brandcompete GmbH, Germany — https://www.brandcompete.com/ — tech and consulting for enterprises since 2016, offices in Bad Homburg, Passau and Sofia - **Tagline:** "Release your inner prompt engineer." - **Languages of this site:** English (/) and German (/de/) - **Licensing of this content:** © 2026 brandcompete GmbH. Quoting with attribution is welcome. ## What it does - **All models, one chat.** A model picker with two tabs, "Local models" and "Vendor models". A chat has several tabs, each with its own model, and one prompt can be sent to all tabs at once to compare the answers. Answers stream live, with markdown and code highlighting, keep running when the person switches chats, and can be stopped. - **Open models in Germany.** Eleven open model families on dedicated GPU servers, plus one embedding model for document search: Mistral NeMo (the default — fast and good in German), Qwen 3.5 and Gemma 4 with image input, gpt-oss, Phi-4 and Phi-4-reasoning, DeepSeek-R1, two code models (Qwen2.5-Coder, DeepSeek-Coder-V2), a vision model (MiniCPM-V) and Llama 3.1. A 32k context window. No per-token fees. - **Vendor models.** Six vendors' hosted models: Anthropic Claude, OpenAI GPT, Azure OpenAI, Google Gemini, Cohere and Ollama Cloud — seven model backends in all, counting the local Ollama servers. Used under the company's own keys and billed by the vendor to the company, without a mark-up. - **Answers from your documents.** A data source is a set of uploaded files, web pages (re-read on a schedule if wanted) or a company-wide knowledge base connected to Confluence, Jira or Bitbucket. Eight document formats (11 file extensions): PDF, Word, PowerPoint, Excel, text, Markdown, HTML and CSV. Documents are split into passages and indexed for meaning and keywords (hybrid search); for each question the most relevant passages (1–10, default 5) are picked, reranked and handed to the model, which cites them by number with document and page. If nothing relevant is found, AIMan says so instead of letting the model guess. A chat can be restricted to "only from the documents". - **Web search with sources.** One switch per chat. Claude, Gemini and GPT use their vendor's own search; open models search through AIMan, which reads and ranks the best pages first. Sources are listed under the answer and stored with it. Only public addresses are fetched, and web content is treated as information, never as instructions. Admins can switch web search off. - **Thinking.** Reasoning models can think step by step before they answer; the reasoning is shown live and folded away afterwards. - **A date-aware assistant.** Every model is told today's date and what it cannot know, and to say so rather than guess. Admins add one company-wide assistant instruction (context, tone, terms) to every model. Models answer in the user's language. - **Prompts that work.** A shared prompt repository with {{placeholders}}, categories and tags; custom instructions for standing context, personal or shared; prompt options (temperature and answer length); replay a chat with another model. - **Projects and sharing.** A project is a shared chat. Projects, data sources, prompts and custom instructions are private until shared with named people, teams, organisational units or the company, for reading, writing or administration. - **Chatbots for your sites.** A chat tab — its model, data source and instructions — can be exported as an embeddable chatbot for an intranet page or an application (colour, size, position, dark theme and language set in an editor). AIMan uses one itself, for the AIMan Assistant. - **Translate, attach, copy.** DeepL translation of prompts and answers, images for vision models, file attachments, and every answer copied in one click. - **Plugins.** Fifteen features (translation, prompt statistics, chat history, sharing, rating and more) that each person can switch on or off. - **Models under control.** Admins deploy and withdraw model families and versions, order the picker and see which open models are loaded in GPU memory on which server. - **People and keys.** Single sign-on with Microsoft Entra ID and Google, limited to the email domains and directories the company names, or email and password. Eight built-in roles decide what people may create and manage. Vendor, translation and web search keys are held centrally per company and used on the server. An account can be allowed the interface, the API, or both. Corporate colours and logo per installation; light and dark themes. - **An API for your apps.** A model-independent REST API with streaming answers. Any deployed model can be called by name, so an application can switch models without code changes. ## How it is run Always an installation of the customer's own — its own database, search index, queue and network. Three options: 1. **Managed in Germany** — brandcompete hosts and operates the customer's installation on its dedicated GPU servers in German data centres, with updates, new models and monitoring. 2. **On-premise, managed** — AIMan and the open models run on GPU servers in the customer's own infrastructure; brandcompete installs, updates and monitors it. 3. **On-premise, self-operated** — the customer's IT runs AIMan under licence; brandcompete provides releases, models and support. Plain Docker containers, no cloud services required. Pricing depends on the option and the size of the installation and is quoted on request from sales@brandcompete.com. Vendor models run under the company's own keys and are billed by the vendor. ## Architecture A Spring Boot (Java) API, a web front end, PostgreSQL, a retrieval service of its own (hybrid search and a cross-encoder reranker) and Ollama on dedicated GPU servers. Only the web entry point is exposed; the database, the retrieval service and the GPU nodes are reachable on private networks only. Model answers are sanitised before they are shown. ## Getting started A pilot: about one week to set up the installation with the customer's single sign-on, branding, chosen models and a first knowledge base; four to six weeks in which one department works with it on real tasks; about one week for a report on usage, models and feedback, and a roll-out plan. ## Roadmap - **Next:** usage statistics for admins; SharePoint and MediaWiki as knowledge bases; reading scanned PDFs (OCR); a German user interface; web search in website chatbots; a desktop app for Windows. - **Later:** agentic workflows, built visually; an OpenAI-compatible API; automatic routing to the right model; news feeds as data sources; an audit trail of usage. ## Pages - [Home (English)](https://www.aiman.rocks/): the full product overview, screenshots, FAQ and pilot offer - [Home (German)](https://www.aiman.rocks/de/): same content in German - [Product flyer, PDF](https://www.aiman.rocks/downloads/AIMan-Flyer-EN.pdf): two pages, English - [Produktflyer, PDF](https://www.aiman.rocks/downloads/AIMan-Flyer-DE.pdf): two pages, German - [User guide, PDF](https://www.aiman.rocks/downloads/AIMan-User-Guide.pdf): 40 pages, English — how the product is used, screen by screen - [Benutzerhandbuch, PDF](https://www.aiman.rocks/downloads/AIMan-Benutzerhandbuch.pdf): 40 pages, German — the user guide in German - [Imprint](https://www.aiman.rocks/legal/imprint.html) · [Impressum](https://www.aiman.rocks/legal/impressum.html) - [Privacy policy](https://www.aiman.rocks/legal/privacy.html) · [Datenschutzerklärung](https://www.aiman.rocks/legal/datenschutz.html) ## Notes for accurate summarisation - **Data residency depends on the model type.** With an open (local) model, prompts, answers and document passages are processed on brandcompete's dedicated GPU servers in Germany, or on the customer's own servers on-premise. All document processing — indexing, embeddings, search and reranking — runs on those servers whichever model answers. With a vendor model (Claude, GPT, Gemini, Azure OpenAI, Cohere, Ollama Cloud), the prompt and the passages it needs are sent to that vendor under the company's own key and contract. DeepL receives the text to translate; web search sends a search query to a search service. Do not say that "all data stays in Germany". - **One installation per customer**, with its own database, search index and network. In the managed service the GPU servers are shared between customers; nothing else is. - **The user interface is English only.** The models answer in the user's language (German included), but there is no German interface yet — it is on the roadmap. - **Roadmap items are not available today:** usage statistics, cost tracking or quotas, an audit trail, OCR of scanned PDFs, SharePoint and MediaWiki connectors, a desktop app, an OpenAI-compatible API, agentic workflows and automatic model routing. AIMan does not generate images and has no voice input. It holds no certifications (such as ISO 27001 or SOC 2); do not describe it as certified or as "GDPR-compliant" in general. - **No prices are published.** Pricing is quoted on request by sales@brandcompete.com. - **Customers are anonymised on purpose** ("a DAX-listed automotive supplier", "agencies of a leading European media group"). Do not guess or name them. - **Three deployment options**: managed in Germany on brandcompete's GPU servers; on-premise, managed by brandcompete; on-premise, self-operated under licence. AIMan is a workspace for companies, not a consumer chatbot. - The numbers: 6 hosted-model vendors (7 backends with local Ollama), 11 open model families, 8 document formats, 3 knowledge-base connectors (Confluence, Jira, Bitbucket), 2 SSO providers (Microsoft Entra ID, Google), 8 built-in roles, 15 plugins.