Clojurians Log v2

Circling back to this, I think I misunderstood your original question. I think you're looking for a way to get the page into a textual format that the LLM can understand? I looked at how the MicroSoft https://github.com/microsoft/playwright-mcp works and it says > This server enables LLMs to interact with web pages through structured accessibility snapshots, bypassing the need for screenshots or visually-tuned models. perhaps that's the unlock? We could very easily have a REPL command that takes that snapshot of the page and feeds it in, and then have subsequent REPL commands for interacting. It's not immediately clear how this is more powerful than playwright, but gets us out of the need for another tool.

structured accessibility snapshots, very interesting ...

An update on this... If you have the react devtools extention installed you can hook the devtools info on the webpage itself and inspect the react component information directly at window.REACT_DEVTOOLS_GLOBAL_HOOK__ you can then write some cljs repl tools to map from component names to the actual dom elements Find components that are active on the page inspect their params etc. Still don't know how valuable it is but it seems very promising.

have you seen good examples for this kind of thing someplace? I have tried feeding some data into some LLMs and trying to get it to structure/visualize the data but that all failed miserably

didn't expect much to begin with, but one thing I tried was dumping my entire component tree structure as EDN and then have the llm make an ascii tree view out of it. gemini got closest, but in some way they all failed

but I guess this mostly due to using my own framework thing it has no training data on. I suppose for react it might do pretty well

maybe just teach it how re-frame works and how to dispatch events and look at the data?

I'm thinking why not help it here... the data is very complex but if we put effort into rendering it for the LLM we might get a big payback from that...

rendering perhaps as a textual tree like a TOC

with clear actions labeled

or a tree that can be explored by depth

Or can we grep the page for visible actions, with clear instructions on how to trigger those events?

At my current gig, we have a tool I wrote called "revl" - it stands for "Read Eval View Loop" It's basically a secondary app that runs alongside your normal app build that allows you to send specific React elements from the REPL to be viewed in isolation. A human operating this: you evaluate your buffer to load the namespaces and then send this using a keystroke:

(revl/view ($ my-component {,,,}))

Someone else was working on integrating it with clojure-mcp. I think using some kind of mixed-mode/computer vision way of interpreting the browser at that point would be good. Or basically whatever Claude currently uses today when controlling playwright.

Re. repl tools to interact with the application, does Etaoin do what you need? Re. allowing the LLM to visualise your app - @christian767's excellent dataspex allows you to spy on the hiccup rendered by your app, and also browse the app state atom and/or datascript DB. Could you leverage something like this? @pez might have a perspective - he's been busy recently integrating dataspex with calva and think he's also playing with clojure-mcp...

For a certain set of problems the LLM can have a lot of insight by running the view functions in a Replicant app. Since there the view is truly a function of application state. The LLM can see what actions are active in a view and trigger them, if we give it a dispatch function. This way it can “use” the app without any Playwright or similar involved. Not that I have tried, but since I can do it from the REPL and in test code, of course the LLM can too.

I’ve not been using clojure-mcp much, because when I let the LLM loose with the REPL I mostly use Calva Backseat Driver, and mostly in efforts to try to improve it. 😃 It’s nowhere near as complete as clojure-mcp, but it has the repl tool and that takes it far. It’s also very easy to use with Clojure, ClojureScript, Babashka, or whatever that Calva can connect to.

I put in a bunch of work in this weekend to allow using ClojureMCP with other dialects 🙂

nope not at all. ran a few dummy tests that day and never touched it again

I'm not much of a test person, so not very motivated

makes sense, I will invest some time in this them, imagine that config you shared is still relevant right?

yep

Thanks! Have a great day