The self-hosted agent fleet that never holds your logins

Grok Bot's whole pitch is that it signs into your apps as you and runs on a shared cloud computer. That is the one thing I designed out of my own agent fleet. NanoClaw does the same class of work on open, self-hosted parts I own, and no agent ever holds a real credential.

Handing an AI your real logins is a large amount of trust to give a vendor that is barely a month old. I already run the alternative: a small fleet of agents on my own servers, built on a lean harness I own and an open-source gateway, where each agent reaches a tool through a scoped key it never sees. Same outcome, a different trust model.

Grok Bot's model: it logs in as you

Grok Bot is a credential-holding agent. It signs into your accounts with your real logins, and all of your Bots share one cloud computer tied to your account, passing files, browser sessions and logins between them with no separate boundary. That is not a side effect, it is the product, and it is what makes it so approachable. I wrote the full review here. The convenience and the risk are the same fact: the shared login.

The open, self-hosted alternative

NanoClaw is the other path. My agents run on my own servers, on a harness I control, and every one of them calls a single gateway instead of any provider or app directly. That gateway is open source (LiteLLM), and it holds every real key in one box. Each agent gets its own virtual key with a budget, so it reaches a model or a data API through a scoped credential it never holds and cannot overspend. No agent sees a real key, and none of them signs in as me. It is the open, self-hostable version of "a team of agents that finishes work", assembled from parts I own rather than rented from a vendor.

The win is control, not price

On raw cost this is not the pitch, because Grok Bot rides on a $30 plan and hosts itself. What the self-hosted path buys is control. The harness is mine, so a patch is a change I choose, not one a vendor ships on a Tuesday that breaks yesterday's work. Every real key, budget and trace sits in one box I own, so a leaked agent leaks nothing and a runaway is capped instead of a surprise bill. Nothing I run holds a human login, and no outside company holds my credentials or my data. Swapping a model provider, or falling back when one is down, is one setting, because the agents only ever know the gateway.

The honest limit

This costs effort, and that is the real trade. I operate the servers, the gateway everything routes through is a single point of failure (recoverable in one setting, repoint an agent straight at a provider), and the roster is still a seed I am watching. Grok Bot asks none of that: you record a task once and it runs. The open path trades that ease for control, so if you cannot run a little infrastructure, it is not the cheaper answer, it is the wrong one.

The bottom line

You do not have to hand an agent your logins to get a team that finishes work. Open, self-hosted parts, a lean harness, an open-source gateway and your own servers, give you the same outcome with the real keys in a box you own and no agent signing in as you. Grok Bot wins on approachability. The self-hosted fleet wins on control, and control is the part I am not willing to rent.