Grok Bot vs Kortix: open source, and a computer per session

The first bot reads the same on both sides

Both products give an agent a computer and a way into your tools, and both describe the result the same way. xAI calls Grok Bot "AI teammates that finish the work": Bots sign in to your tools, use them like a person would, and come back with finished work (x.ai/bot). Kortix is the open-source CompanyOS, and an agent there is instructions, a model and permissions, held as files in a repo your company keeps.

A reader comparing the two on features alone finds a short list. The two products part company on the second bot, or the second person, or the first client who asks who approved something. Those are the moments a shared machine and a settings screen stop being enough.

One computer, two Bots, one browser

xAI is direct about the machine. "Every Grok Bot shares one persistent cloud computer. Your Bots share that machine (files, browser, logins), so they can hand work off and keep context. Isolation is per user, not per Grok Bot" (x.ai/bot). Shared state is the feature. Two Bots of one user pass a project between themselves, and neither one starts over.

Kortix moves the unit down a level: one isolated sandbox per session, each session with its own isolated machine and branch. The Kortix security page phrases it "One cloud computer per session", with "Its own branch of the repository". The handoff happens through the project and its change requests, not through a shared browser profile. Two sessions working at once share no file tree, no cookie jar and no logged-in tab.

The difference stops being philosophical the first time two clients land in the same account. A studio that runs both retainers through one bot machine has both sets of logins in one browser. A session with its own computer cannot post as the wrong client, because the other client's session was never there.

What an approval covers, and what it leaves behind

Grok Bot reviews each action before it runs, and its documentation is careful about what that buys: "An approval controls the proposed action. It does not reverse work already completed" (Grok Bot docs). The card appears before the send, the command or the purchase, which is the right moment to catch a mistake. A card is the wrong place to answer a question asked a week later about what the bot did on Tuesday.

Kortix puts the gate one level up, on the change rather than the click: session work reaches main through a change request, and merge is default-deny for agents (Kortix security page). A reviewer reads the diff, sees which agent opened it and what it touched, and merges it only if that reviewer holds the merge capability. What was approved is therefore a commit rather than a card that closed.

The controls a regulated buyer asks about next split along the same line. The organization-wide enable switch, network controls, action recording, enforced auto-review, audit logs, OpenTelemetry export, an MCP allowlist and SCIM are enterprise-only in Grok Bot, and "self-serve Teams do not see those settings" (docs.x.ai/grok-bot/security). Kortix runs on Kortix Cloud, in your own VPC, or on your own on-prem network, with any model and your own keys, so the answer to where the work happens is a deployment choice rather than a tier.

Where the rules live after the trial

In Grok Bot the rules are settings in the x.ai apps: Auto Review on or off, and custom rules that allow or ask, set by the person who works there rather than by the Bot. Team rules arrive with the Enterprise plan and sit in the same table, marked as required by an admin.

In Kortix the rules are files. An agent's instructions, the model it thinks with and the permissions it holds sit in the project repo, next to the skills and the connector grants, and a change to any of them arrives as a diff a reviewer can read. Nothing about that is exotic; it is what makes the second year of an agent program an edit instead of a migration.

What running Kortix looks like

$ curl -fsSL https://kortix.com/install | bash$ kortix init my-agent$ kortix ship$ kortix sessions new$ kortix cr merge

The install line fetches the CLI, kortix init writes the agent into the project, and kortix ship sends it. Sessions then run on their own computers, and kortix cr merge is where a change request closes, by whoever holds the merge capability. The commands come from the Kortix host docs.

The question to ask before either invoice

Ask who has to answer for the work three months from now. If the answer is the person who ran it, on their own laptop, a shared machine and a plan already paid for are enough, and OpenClaw covers that week at no subscription, as does self-hosted OpenMausBot, which also sells a hosted Starter plan at $15 a month.

If the answer is a team, a client or an auditor, two things decide it: a computer the rest of the work cannot reach, and a record of what changed and who merged it. Kortix is built for the second answer. The self-hosting route covers deployment, Grok Bot security covers the controls, and Grok Bot pricing covers the arithmetic of a seat against a credit.