Asteria Cloud — Enterprise agentic AI, sovereign and sourced

The problem

An organisation’s critical knowledge is scattered: technical documentation, specifications, test reports, spreadsheets, procedures. Finding it and turning it into something useful costs your most qualified people their most expensive hours. Consumer AI tools, meanwhile, know nothing about your data — and giving them access would be a serious confidentiality risk.

What Asteria Cloud is

Asteria Cloud is an agentic AI platform for organisations. A central conversational agent talks with the user, searches the documents and collections you give it, runs analysis code in an isolated sandbox, and delegates specialised work to sub-agents. Every answer cites its sources. Every agent acts strictly within the permissions of the user who invoked it, and consequential actions stop to ask for human approval before they happen.

Four pillars

Pillar 1

Model-agnostic hub and sovereignty

Many models, no vendor lock-in: Use the strongest model for each task — Gemini, Claude, GPT, Mistral, or one you host yourself — and switch without rewriting your agents.
European hosting: Scaleway infrastructure in France, or deployment on your own. Your provider keys stay yours.
Secure by design: Strict per-user and per-organisation isolation, encrypted secrets, tamper-evident audit logs, and controlled outbound network access.
Native data analysis: Sandboxed code execution for querying complex datasets and producing charts, with no path to the host environment.

Pillar 2

Document and context engine

Talk to any format: Upload PDFs, Excel and CSV spreadsheets, images. Content is extracted, indexed and immediately answerable in natural language.
Hybrid retrieval: Vector, keyword and entity-graph search combined over large technical corpora. The agent answers from your organisation’s truth, not the web’s.
Sourced answers: Text, tables and charts come with exact citations pointing back to the passage they came from. An unsourced claim is visible as one.

Pillar 3

Agents, skills and workflows

Reusable agents and skills: Your teams compose their own agents, and their own skills — packages of know-how loaded on demand — to automate the work that repeats.
Human approval on consequential actions: Autonomy is tunable. An action with consequences pauses and asks for explicit approval, stating precisely what it is about to do.
Connected to your systems: Enterprise connectors reach into your existing systems to break data silos, with credentials owned per organisation and revocable at any moment.
Durable workflows: Chain agents and tools into repeatable processes, run on demand or on a schedule, surviving restarts and pausing at approval points.

Pillar 4

Governance and developer platform

Central administration: Manage models, users, permissions and storage quotas from one console, with usage tracked per team.
Sharing takes consent: Shared workspaces and collections rest on explicit contributor consent — nothing becomes visible to colleagues by default.
Multilingual by default: Interface in English, French, Spanish, German and Italian, so distributed teams work in one place.
OpenAI-compatible API: Organise work into projects, issue API keys, track usage, and call the engine from your own tools through an OpenAI-compatible endpoint.