KohakuTerrarium

The deciding trait is that you stop rebuilding the same plumbing. Controller loop, tool dispatch, triggers, sessions and persistence already exist here, so your third agent idea costs a config file instead of another repository you will have to maintain. The bundled SWE creature means you can watch a working coding agent before you write anything of your own.

You will find this too heavy if you only ever want one small script. Pick it when you expect to build several agents. Pick LangGraph if you would rather use the framework everyone else has already debugged.

The risk is blast radius. A shell tool and file editing ship in the box, the sandbox does not, and no git operations are wired in either. So a creature that misreads its instructions writes to your working tree directly, with no container between it and the rest of the disk and no commit boundary to roll back to.

Hot-plug makes this worse: topology changes while the engine is running, so the thing that misbehaved may no longer be attached when you go looking for it. What it does right is naming its tools individually, so you can see what a creature was given.

1. The abstraction boundary is stated rather than implied: a creature is self-contained and the hosting engine owns channels, lifecycle and topology bookkeeping. That separation means a failure can be attributed to one side or the other, which is more than most orchestration layers permit.

2. Composition is expressed as a graph, so the shape of a system is a declared artefact and not an emergent property of prompt text.

3. No evaluation accompanies the design, and none is claimed. An architectural argument published as an architectural argument is the honest form, though it leaves the boundary's value unmeasured.

Kohaku Lab publishes this with no hosted tier and no commercial edition, which means there is no revenue line to protect and nothing here for an acquirer to buy except attention. Just under five hundred stars is interest, not adoption, and pays no salaries.

Moat: none of the durable kinds, though the substrate framing is a real position if enough builders standardise on it. Likely path: a maintainer's side project that either finds a company to fund it or slows to occasional releases. Position: build on it if you would be comfortable maintaining the fork yourself.

Nothing per seat across sixty desks, and inference billed to accounts we already reconcile, so finance is a formality. The Studio layer is the only part that looks like an operations surface: a catalog, identity and sessions, running on a server we would have to stand up and patch ourselves. There is no vendor to call when it stops.

It does not run unattended, so it never becomes a step in a pipeline I can measure or gate a merge on. Onboarding costs a Python developer a week of reading. Not yet: revisit when there is an unattended runner and a documented retention policy.

The licence is the interesting part. KohakuTerrarium 1.0 is Apache-2.0 with a name tax: fork it and your fork has to carry Kohaku or Terrarium in its name and link back visibly. That is honest and it is in the file, but read it before you plan a fork.

Everything else is mine. Keys go in with kt config key set, MCP servers attach over stdio or streamable HTTP and can be scoped to one creature instead of the whole process, and the notebook and search tools are already there rather than hunted down as plugins. No local serving option is documented, which is the one gap.