Cheshire Cat AI

The trait that decides it is that you can see the machinery working. It runs as a service with an admin interface and an interactive API alongside it, so poking the agent, watching what the hooks do and breaking it deliberately is the intended activity rather than an advanced topic.

That is genuinely the best way to learn this material and a poor way to ship anything, because the current version is explicitly unfinished. Pick it to understand the shape of the problem. Pick Flowise when you need a result rather than an education.

The maintainers have written the warning themselves, which spares me the argument and leaves the second problem. Plugins can add custom endpoints to the running service, so the API surface of a deployment is whatever the installed plugins decided, and nothing documented constrains what an endpoint may do or requires it to be authenticated.

What it does right is saying so. A project that marks its own release unstable is more useful than one that ships the same code with a launch post, and the honesty is worth a point on its own.

1. Exposing the loop over HTTP with a browsable specification is pedagogically strong: a student can call each part in isolation and watch the state change, which a library import does not permit. 2. Hooks and directives give named extension points, so the places where behaviour can be altered are enumerable rather than discovered.

2. No evaluation exists and none is claimed, consistent with a project describing itself as a teaching and research framework rather than a performance one.

The model is honest and it is a services model. Revenue comes from a maintainer's time rather than from a product, which makes the project durable in proportion to how much teaching work it attracts and caps growth at the number of hours in a week. Copyleft licensing is consistent with that, because it discourages a competitor from hosting a proprietary version.

Moat: the author's reputation in a specific community. Likely path: steady existence as an education asset. Position: fine to learn on, not something to underwrite a product with, and the support you buy is a person rather than a company.

Zero cost, and two blockers that both take weeks. The licence is strongly copyleft, which is fine internally and becomes a legal conversation the moment anything derived from it reaches a customer, and my counsel will want that conversation before the pilot rather than after.

The second is the maturity label, which makes the vendor's own recommendation the reason to decline. Nothing runs unattended, so it produces no scheduled work I could measure the risk against. Not yet: revisit when the authors call a stable release.

Strong copyleft on a framework is a choice I respect: whoever builds on this has to publish, which is the opposite of the permissive-then-proprietary pattern I keep watching. The install is one command with a modern Python tool rather than a page of prerequisites.

Plugins register tools, hooks and directives, so extending it means writing a module rather than forking the core, and the core is small enough that I would read it anyway. No protocol client, so my existing servers stay outside, and that is the main thing I would add first.