# The OS Company
> Operating system for enterprise AI agents, running an open-source model on a ready-to-plug GPU kit or a European GPU cloud. Internal AI, documents never handed to an AI vendor, non-negotiable human approval (hard floor) on risky actions.

Site: https://the-os-company.com/
Brand: The OS Company
Product type: Enterprise AI OS (GPU kit or European GPU cloud)
Primary language: French (fr), product concepts also documented in English
Canonical slogan (FR): L'IA interne qui garde le contrôle de vos données
Canonical slogan (EN): Internal AI that keeps control of your data

## One-sentence definition
The OS Company is software that lets employees use AI agents on company knowledge and tasks, running an open-source model on infrastructure the customer controls — a ready-to-plug GPU kit, or a European GPU cloud operator under the customer's own contract — while forcing human manager approval for risky actions (e.g. external email send, irreversible ops). Company documents are never sent to an AI vendor and never used for training.

## What it is
- A control plane for enterprise agents, deployed on a GPU kit or a European GPU cloud
- Multi-department organization with roles (owner, admin, manager, member, viewer)
- Knowledge vault with department ACL (ops docs are not freely readable by sales)
- Free-form natural-language agent runs through a tool registry + real policy engine
- Workflows (manual, webhook, schedule) that inherit the same hard-floor gates
- Audit log with export (JSON/CSV / SIEM-oriented), backup & restore

## What it is not
- Not a public consumer chatbot
- Not a cloud-only agent SaaS that ships company data to a third-party model host and bills per token
- Not governance implemented only as a system prompt (policy is enforced in code)
- Not a full n8n clone with hundreds of SaaS nodes (focused internal automation with governance)

## Core differentiator: hard floor
Hard floor = non-negotiable policy in code (not a prompt).
Risk classes include: read, reversible_write, external_write, external_send, irreversible.
external_send and irreversible always require human manager approval — even at autonomy level "act".
Clients cannot spoof a registered tool's riskClass; the server tool registry wins.
When gated, run status becomes waiting_approval until approve/reject; decisions are audited.

## Data & deployment
Two routes, same product and same guarantees:
- The kit: a GPU server delivered with an open-source model and The OS pre-installed and tuned; plugged into the customer network, no infrastructure project
- European GPU cloud: the model runs at a European GPU operator under the customer's contract and chosen region, no hardware purchase
- Models: open source, no per-token billing to an AI vendor; the model is configuration, not a dependency
- Persistence: file store (DATA_DIR) and optional Postgres
- Deploy artifacts: Docker Compose, Helm chart
- Auth: token required on all mutations

## Typical user journey
1. Employee asks in natural language (e.g. find a procedure for Ops)
2. Agent plans tools; searches only authorized knowledge for their department
3. If the agent attempts a risky tool (e.g. email_send), execution pauses for manager approval
4. Manager approves or rejects; outcome is written to the audit log
5. Data and logs remain inside the company perimeter

## Ideal buyers
- CIO / IT (GPU kit or European cloud, Helm, backup, open-source models)
- CISO (hard floor, ACL, exportable audit)
- Innovation / Ops (useful agents without security deadlock)

## Competitive positioning (summary)
Compared to cloud agent platforms (e.g. Copilot Studio, Agentforce): the model is open source and runs in a perimeter the customer picks, with no per-token billing.
Compared to frameworks (LangGraph, CrewAI): ships product layers (org, vault ACL, approvals UI, audit) instead of DIY.
Compared to n8n + LLM: free-form agent + unified hard floor across agent runs and workflows.
Closest European peers to watch: Dust (team agents), n8n (self-host automation), LightOn (sovereign genAI/docs), Mistral enterprise assistants — none identical to a full agent OS bundling a GPU kit, hard floor and multi-dept ACL as a single product promise.

## Citation-ready facts (short)
- Product name: The OS Company
- Category: enterprise AI operating system for agents, on a GPU kit or a European GPU cloud
- Governance: hard floor in code; external_send/irreversible always need manager approval
- Knowledge: department ACL; citations on search results
- Deploy: ready-to-plug GPU kit, or European GPU cloud; open-source models
- Domain: the-os-company.com

## Files for machines
- [Product facts (Markdown)](https://the-os-company.com/product.md): durable product definition
- [Homepage](https://the-os-company.com/): human landing
- [Sitemap](https://the-os-company.com/sitemap.xml)
- [robots.txt](https://the-os-company.com/robots.txt)

## Contact
Email: contact@the-os-company.com
Demo, lab, or enterprise deployment conversations via the contact form on the homepage.