# Open-source agents for one person, or for a whole company

> OpenClaw handles one person's inbox, calendar and travel well, from a chat app on a machine you own; Kortix is the open-source alternative for a whole team. Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers.

URL: https://companyrunbyagents.com/kortix-vs-openclaw

## Where the two part ways

One runs on a machine you own; the other is a project a team owns.

| | Kortix (recommended) | OpenClaw |
| --- | --- | --- |
| What it is | A company's agents, skills and memory in one repo | An assistant that does small jobs from chat apps |
| Where it runs | Kortix Cloud, your VPC, or on-prem | On the machine you install it on |
| How a team works in it | Thousands of agents on one config, each on its own computer | Run it yourself or share one install with the team |
| How a change is approved | Session work reaches main through a change request | The owner approves the big calls by message |
| Memory | Company memory files in the project repo | Markdown files in each agent's workspace |
| Self-hosting | Cloud, VPC or on-prem; self-host is free | No hosted tier; it runs on your machine |
| Open source | Yes | Yes |

## What each one is

OpenClaw calls itself open source and runs on the machine you install it on. No subscription, no hosted tier. Its homepage lists what it handles: an inbox organized, an email sent, a calendar kept, a flight checked in, all from WhatsApp, Telegram or another chat app. It installs on macOS, Linux and Windows. An independent 501(c)(3) stewards the project.

Kortix is built the other way round. The unit is a project, and a project is one git repository. Its agents, skills, memory and connector settings are files in it. `kortix.yaml` at the root names the agents, what each one may reach and what starts it. [Company as code](/company-as-code) is the long version of that idea.

Both products are open source. Past that line they answer different questions. One is a program a person buys a machine for. The other is where a company keeps the rules its agents follow. Day one asks two different questions: can I run it, and who else can read what it knows.

## Where the agents run

A company-size deployment is assembled by hand, and one owner published the assembly: seven agents on a single 4GB Hetzner box, a CEO, a CTO and five specialists ([write-up](https://hellonunc.com/blog/posts/company-run-by-ai-agents/)). The two main agents ran as separate Linux accounts, the CEO on port 18789 and the CTO on 18790, each with its own gateway and a systemd unit set to survive a reboot, reached over Tailscale. A cron job polled both health endpoints every five minutes and restarted the silent one. In the first week that caught one crash by itself; the owner fixed the other.

That is a working company on modest hardware. The maintenance shows in the same text: Linux accounts, ports, service units, SSH keys, health scripts. A person assembled it. A person keeps it running. A single failed box takes the org chart with it, which is why the write-up ends up describing mutual supervision between agents.

Kortix moves that layer behind the product. A session runs on its own cloud computer, with the project's repository checked out and a branch of its own. Agents live in one config, not in a set of Linux accounts. The tenth agent is a block in `kortix.yaml` and a review, not another port and service unit on somebody's server. Thousands of agents in parallel on one config, each on its own cloud computer.

Self-hosting suits both, in different shapes. OpenClaw has nowhere else to run, since its own page says there is no hosted tier. Kortix is the one with a choice of homes. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

## Who approves a change

In the [same account](https://hellonunc.com/blog/posts/company-run-by-ai-agents/), the approval surface was the chat window. The owner sent instructions from a laptop over SSH, approved the large calls, and read a status summary about once a day. Nothing in that loop produced a diff, so each check-in weighed the agent's account of the week against the owner's own memory of it. Week one ended with a product and 158 tests. It also ended with half the agents' own files out of date and two of the seven still unnamed. The check-in happened in chat, not in those files.

Kortix puts the review somewhere a reviewer can open. Session work reaches `main` through a change request. Merge is default-deny for agents. An agent that edits the project itself, a skill, a prompt or a line of memory, opens a change request. A person reads the diff on the Review page and merges it. The unit of review is a change to the company. Most of them are a few lines: a skill that gained a tool, a memory file that gained a correction, a rule that now asks first. Small enough to finish with coffee, specific enough to argue with.

For one owner, chat approval costs nothing and needs no repository. The cost arrives with the second person. They have to ask what an agent may do, and the answer comes from someone's recollection. A diff answers it once, for everyone who reads it later.

## Which one fits which team

Take OpenClaw if the work belongs to one person and lives on a machine that person controls. Its advertised jobs are personal: the inbox, the calendar, the flight. A single owner who wants an assistant in WhatsApp, no subscription and no hosted account gets exactly that. The [seven-agent account](https://hellonunc.com/blog/posts/company-run-by-ai-agents/) shows it scales to a small hierarchy when someone is willing to run the server.

Take Kortix if the agents belong to a company. The agents, their skills and the company memory sit in one repository the team shares, and a change to any of them arrives for review. A new colleague reads the agents, the rules and the memory in the same place the review happens, and the version that ran last quarter is a commit.

Some teams end up with both. One person's assistant handles that person's inbox, while the company's agents work from the repository. Deciding which work belongs on which side is the useful part of the choice, and a team can revisit it each quarter.

For a team, begin with the repo. Kortix is the recommendation, and the reason is scope rather than capability. The work a company needs reviewed has a home in a project. A personal assistant on a laptop leaves that home empty. The usual order is a company system first, then assistants for the people who want one.

## The review rule, in the manifest

`kortix.yaml`

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

Source: https://kortix.com

Two connectors from the project shown on kortix.com, with the rule that holds one kind of call for a person.

## Four questions this comparison turns on

Short answers, in the order buyers ask them.

### Which alternative fits a team?

Kortix, when the agents belong to a team. It keeps the company's agents, memory and permissions in one git repo that a person reviews, where OpenClaw puts an assistant on each person's own machine.

### How is OpenClaw different from other agents?

It runs on your own machine and is used from chat apps such as WhatsApp and Telegram, with no subscription and no hosted tier. Its memory, skills and models stay on hardware you own.

### Can OpenClaw run a whole company?

Yes, one owner can run a small company on it, and a [published account](https://hellonunc.com/blog/posts/company-run-by-ai-agents/) documents seven agents on a single 4GB VPS. The limits are operational: one machine, one gateway per agent, and a person reading the results.

### Is Kortix free to self-host?

Yes. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

## Put the company's agents where a person can review them

Start with one agent block in a repo you own, then add the skills, connectors and rules as you read them.

[Try Kortix](https://kortix.com) · [Read the docs](https://kortix.com/docs/developers)