GoClaw

The deciding trait is that there is a small version. A desktop build runs on a local file database with up to five agents, so you can find out whether you want this before agreeing to operate anything. Most projects in this shape make you stand up the full server first and decide afterwards.

Be clear about what it is: a place for assistants that talk to you, remember things and hand work to each other, rather than a tool that rewrites your codebase. Pick the small edition if that is the itch. Pick a coding agent if you were expecting diffs.

The cost of this design is operational and it is permanent. Working, episodic and semantic memory all accumulate, backed by a vector store inside a relational database somebody is responsible for, and no retention policy, no eviction rule and no size ceiling appear anywhere in the documentation.

That is a disk that only grows, holding whatever an assistant decided was worth remembering from every conversation it ever had. Nobody notices until the volume fills. What it does right is separating the tiers by name, so at least the thing that grows is legible when somebody finally goes looking for it.

1. Naming the stages has a consequence: a failure can be attributed. When an answer is wrong the question becomes which stage produced the error, and that is answerable in a way it is not for a single opaque call. 2. The ordering is conventional and none the worse for being so.

2. No evaluation accompanies the claim, and this is a case where one would be straightforward, since an eight-stage loop can be ablated a stage at a time. That is exactly the experiment the design invites and nobody has run it in public. Legible architecture, asserted benefit.

The product is a rewrite of an existing design in a different language, which says where the value is claimed to sit: speed and packaging rather than the idea itself. That is a genuine advantage and not a defensible one, because the next rewrite is somebody else's weekend and the idea is already public.

Three and a half thousand stars proves appetite for the category and says nothing about willingness to pay. Likely path is a hosted service, which is the only thing this shape can charge for. Position: interesting to watch, and I would not put a company workflow on terms nobody has tested yet.

Somebody here has met a security team. Roles, isolated per-user workspaces and encrypted credentials at rest are the three things I ask about first, and they are in the architecture rather than on a roadmap. That is unusual enough that I read the rest of the documentation carefully instead of skimming it.

The rest is where it stops. No identity provider integration, so those roles are maintained by hand across sixty engineers, and no audit export, so I can demonstrate nothing to a regulator. Not yet: I need single sign-on before roles somebody edits by hand count as access control.

This is my specific complaint about the whole category. The code is right there, the binary is one file with no runtime to install, and the terms say non-commercial, so the thing I can most easily read is the thing I can least freely use. A licence like that makes a fork legally pointless.

The rest is what I want. Twenty-odd providers, any OpenAI-compatible endpoint so the weights can be mine, and MCP servers that attach as tools. Grudging respect for the engineering and none at all for the clause, which turns an otherwise excellent piece of work into something I cannot recommend.