On Apr. 19th, 2025, we had the fourth meeting of scicloj-ai-meetups, a new Clojure dev group focusing on AI, continuing the earlier scicloj-llm-meetups group.
17 people attended the meeting.
Brief summary
The speakers may add more information at the comments of this page.
Baruch Berger - Weaving LLM tools into the feedback loop
Baruch gave an overview of various workflows and practices using LLMs combined with other tools. In particular, Baruch presented a tree-based canvas UI for interacting with LLM conversations.
theronic - Why MCP Sucks & How To Use It Anyway @theronic discussed the concept of MCP (Model Context Protocol), and presented Modex, an MCP library in Clojure. The talk also demonstrated how to create an MCP server and also discussed the Datomic MCP Server.
Recording
The Scicloj-AI-meetups group
The scicloj-ai-meetups group is a new developer community focused on AI models, their underlying principles, and practical applications.
The group is organized by @stoica94 and @daslu, emerging from community discussions about the need for more open collaboration and experimentation in AI. It builds upon our previous initiatives like the scicloj-llm-meetups, but with an expanded scope that explores diverse types of AI models beyond just LLMs.
While some content will be specific to the Clojure ecosystem, much of the material will be valuable and relevant to the broader AI and programming communities.
That was a very productive meeting. Thanks for putting it together Daniel.
Both were great presentations. So thanks to Baruch and Petrus for sharing.
I did take a few notes, if anyone finds them useful.
# LLM Conversation Canvas (not the real name) (Baruch)
Is it your tools that handles "agent loop" until an accetable answer (w/ tools: MCP (ex Manus w/ manager agent, sub-agents))? If so, how do you determine the correct answer?
. That tool is using Claude to iterate until the correct answer.
. Claude has the notion of stop tokens... manually interject, when to stop iterating
! Claude MCP (can store memory in a graph)
It looks the tool can go backwards and forwards in the conversation. Is that right? (... didn't get to ask)
Which executable are you using for the "screenshot" tool? (... didn't get to ask)
Can you slot in an external editor (Emacs)? (... didn't get to ask)
https://github.com/modelcontextprotocol/servers/tree/main/src/memory
# MCP server (Petrus)
Are there "open" alternatives to MCP... as an interface between multiple agents
ex: for "planning ahead"... mentioned PDDL, which looks like a DSL?
"tool" calls
Google: phoenix: google/A2A
OpenAI: I think OpenAI has its own "proprietary" Agents API.
... crux of the issue seems to be if/how to "lock down" how LLMs coordinate
"planning ahead"
"tools"
"security" concern
"discoverability" of tools
https://petrustheron.com
https://github.com/theronic?tab=repositories&q=prompt&type=&language=&sort=
https://github.com/theronic/modex
https://github.com/theronic/clojure-prompts
Google: PDDL ("The Planning Domain Definition Language (PDDL) is an attempt to standardize Artificial Intelligence (AI) planning languages.")
Google: Agent MCP Alternatives (this is changing from hour to hour, lol!)
https://mcpservers.org/
https://github.com/google/A2A - An open protocol enabling communication and interoperability between opaque agentic applications (MCP alternative)
