# Open-source private AI for business, kept as files

> Private AI for business comes down to where the agent's rules, memory and history live, as much as where the model runs. Kortix is the open-source AI Operating System. In it, every agent, skill, rule, connector and piece of context is a file in one git repo.

URL: https://privateaiforcompanies.com/

For teams that keep the rules, the memory and the record

**The open-source AI Operating System**

`kortix.yaml`

```yaml
connectors:
  - slug: linkedin
    provider: composio
    policies:
      - match: "linkedin.create_*"
        action: require_approval
```

Source: https://kortix.com/

Any model provider with your own keys. Or the ChatGPT plan you already pay for.

[Try Kortix](https://kortix.com) · [What self-hosting gives you](https://privateaiforcompanies.com/self-hosting-kortix)

## Three shapes of private AI, and what each one rents

The model is one decision; the rules, the memory and the record are the other three

### Model in your VPC: The model runs inside the perimeter. The working knowledge does not.

Inference stays on your network, while the prompts, the rules and the history that make the model useful sit in the vendor's product.

What the review found: `acme-private-ai-review.md`

- Model endpoint: your VPC
- Agent definitions: vendor console
- Session history: vendor database
- Export path: support ticket

Acme's deployment review, one page

### Vendor workspace: Encrypted, and still their workspace.

A hosted private workspace encrypts what it stores and keeps the SOPs, the skills and the run history inside its own product.

What the review found: `northwind-workspace-review.md`

- Encryption at rest: yes
- SOPs edited in: their builder
- Skills and prompts: their library
- Run history: their database

Northwind's procurement note, one page

### Self-hosted: The rules are files you can diff.

Self-host Kortix with one Docker Compose stack and the agents, skills, memory and connectors come as files in a git repository on your server, with the approval rules in the same file.

What the review found: `kortix.yaml`

```diff
connectors:
- slug: linkedin
+    policies:
+      - match: "linkedin.create_*"
+        action: require_approval
```

+3 −0

One change request, reviewed by a person

## What each shape leaves in your hands

Kortix, Prem AI and VDF AI, on the six questions a private AI decision turns on

| | Kortix (recommended) | Prem AI (Fluso) | VDF AI |
| --- | --- | --- | --- |
| Open source | Yes. Read it, fork it, audit it. | Not published. | Not published. |
| Self-hosting options | Kortix Cloud, your VPC, or on-prem. Self-host is free. | Not published. | On-prem, private cloud, sovereign air-gapped. |
| Where agent work runs | Sessions run on Platinum sandboxes, not on the server. | On Prem's infrastructure, attestation at every layer. | Agents, language models, retrieval and orchestration run inside your own infrastructure. |
| Where the rules live | In kortix.yaml, in the repo you own. | In Fluso's no-code agent builder. | In VDF's skills and workflow editors. |
| Approvals and audit | Approval gates you set. Off until you set them. | Not published. | Access controls, execution logs and audit trails. |
| Pricing model | Free $0, 500 credits per month; credits from $50 a month. | No public price; a demo request. | On-prem annual seat bundles; cloud plans from $20. |

Each rival cell is what that vendor's own pages say, read 29 October 2026. Where the limit sits: agent sandboxes run on the compute provider you configure, and the instance stays reachable so they can call back, so the isolated machine is rented even when the operating system is not.

## What stays on your disk

The three answers a procurement review asks for first

- Data on disk When you self-host, the database, the files, the repos, the secrets and the audit record all sit on disk you control.
- Approval gates Approval gates you set. Off until you set them.
- Audit record Every action is recorded. Reading, exporting and streaming the audit log depend on the plan.

## What buyers ask before a private deployment

Short answers, with the limits named

### Is there an AI that is completely private?

Not entirely. Every agent runs somewhere, so private means owning the files and the record while the compute stays rented. When you self-host Kortix, the database, the files, the repos, the secrets and the audit record sit on your disk; the sandboxes run on the compute provider you configure.

### What is the 30% rule in AI?

A rule of thumb with no legal force: AI takes about 70% of repetitive, data-heavy work and people keep 30% for judgement and approval ([Intersog's definition](https://intersog.co.il/blog/what-is-the-30-rule-in-ai/)). Private AI only holds that split when the approval step and the record sit in the system that runs the agents.

### Which AI is best for a small business owner?

The one whose files, memory and bills you can take away. Start on the Free plan, $0 with 500 credits per month on your own provider keys. Self-host is free.

### How much does a private AI cost?

It depends on what you rent. Self-hosting the Kortix platform is free; you pay for the server, the model calls and the sandboxes. On Kortix Cloud, Free is $0 with 500 credits per month and paid credits start at $50 a month.

## Own the operating system, rent only the compute

Run Kortix on its cloud, in your VPC, or on your own server, and keep the files either way.

[Try Kortix](https://kortix.com) · [Read the self-hosting docs](https://privateaiforcompanies.com/self-hosting-kortix)