LocalAGI

The deciding trait is that there are no keys to obtain. This is built to run against models on your own machine, with nothing to sign up for and nothing to configure at a provider, which makes the first hour of using it unusually short and unusually quiet.

What you get in exchange is exactly as good as your hardware, and somebody should say that out loud: a consumer graphics card is not a frontier model, and no interface fixes that. Pick it if privacy is the actual requirement. Pick a hosted assistant if quality is.

The risk is an open interface with no stated gate. Making each agent look like a familiar API is a genuinely useful decision, and it also means anything able to reach the port can instruct an agent that has memory, tools and a knowledge base behind it.

On a laptop that is fine. On the machine in the corner that everyone's containers can reach, it is a service nobody registered, and no authentication scheme or access rule is described anywhere. What it does right is choosing a shape people already know, which is why anybody will find it usable at all.

1. The claim is doing real work. Turning one instruction into a group of cooperating agents requires decomposition, assignment and a termination rule, and none of the three is described. 2. What is documented instead is the surface: agents exist, teams exist, a prompt produces them.

2. This is the difference between a capability that is demonstrated and one that is asserted, and the distinction matters most exactly where the claim is most attractive. No evaluation is published. A reader who wants to know how well it works has only the interface to look at.

The dependency defines the risk. This exists to sit beside a local inference server maintained by the same person, which is efficient and concentrates the failure mode: one author's attention supports both, and for most users the load-bearing half is the other one.

Three years of releases is a real signal and the best available here, since consistency over time is the one kind of maintenance evidence that cannot be faked. Moat: none, and none needed by a project with no revenue to defend. Position: use it, and watch the companion project rather than this one.

The install is the easiest part of my week: one command and a service my platform team runs wherever we already run things. Scheduled tasks are a genuine operational feature, because an assistant that acts on a timer is one I can put on a runbook rather than in somebody's browser tab.

Everything above that is missing. No accounts, no roles, no retention policy, and nothing producing a record of which agent did what for whom. Sixty engineers cost nothing and govern themselves, which is another way of saying nobody governs them. Not yet: identity first, then a pilot team.

Scripted actions in a real language is the feature I keep asking for. Not a configuration dialect with three conditionals, an actual interpreted language, which means what I write to extend an agent is what I would have written anyway and I can test it on its own.

Skills live in a directory format that syncs from a repository, so my agent configuration is version-controlled like everything else instead of clicked into a database. The licence is permissive, the whole system runs on my own hardware, and MCP servers attach. There is nothing here I am renting.