@Jay Milne when using the MCP feature that is built in it sends ALL instructions for EVERY MCP server that you have installed in basically every prompt, (to my understanding and from what I can see). This significantly bloats the context window and leads to more context rot and higher token count therefore higher cost (if you are using a local or free model). Either way the context rot is always an issue. Just installing 5 to 7 MCP servers can be upwards of 50k tokens in directions that are sent in the context). How mcp-cli helps is that you actually install the servers into mcp-cli, then ONLY the directions for mcp-cli are forwarded to the model (roughly 400 tokens). I simply gave the mcp-cli github to A0, told it i wanted to upgrade our mcp calling framework and then asked it what mcp servers it recommended to install to test. It suggested 7 seperate servers, found the code for all of them, installed them into mcp-cli, updated the system prompt and tested each function within each MCP it installed before it installed the next. Now using this method you could install virtually limitless mcp servers and not have to worry about any token bloating, mcp-cli has a built in search so it only calls the tools that are relevant to that particular query.

@Joshua Hupp right out of the box, it installed the cli, updated the system prompt to direct all mcp functions to mcp-cli itself and then installed the 5 most important mcp servers that it thought it needed most "i prompted to do this" just testing it out. After installed i prompted to test all tools and functions within each server it installed individually and no issues, i have installed roughly 13 or so, all work no problem yet. I didn't touch any of the agent framework just simply bypassed for a better option. I have found that when trying to upgrade A0 through A0 to have a conversation and suggest what you would like to implement and build a plan with the agent on how to best apply it using best practices and keeping the A0 core operation in mind. I am also doing this with a thoroughly enhanced RAG system. Not bypassing the A0 memory framework, just building a later on top of it for deeper knowledge ingestion. I love A0 the vanilla framework is amazing, also gives you a lot of control to add or tweak things to your particular use case. In 5 days it has become my daily driver. I figure it will take me about a week to get everything I have planned for it incorporated. Plan is to build it all out first, then the knowledge ingestion. I have it linked to a google drive, notebooklm, obsidian, and Perplexity. All fresh instances. If all works as it should my entire knowledge base should finally he completely organized, synced and searchable between my entire toolchest. We will see in about a week, fingers crossed but so far so good.

@Frank Hurt i barely use n8n anymore but i do have some workflows setup that i use from a while back, just really looking to turn A0 into my all encompassing daily driver. I'm sure there are some kinks but for what i need it was simple and works. I will keep everyone updated on my additions.

I am pretty sure you can use your claude subscription via claude code. Just verify, that all coding task are not to use the code execution tool but use claude code. Have seen this done jordanurbs i know did this, or at least attempted to. Not sure on GPT codex, and for cursor, I know that cursor has a web app, you may be able to ask the agent to use the cursor webapp but being that it is not a command line tool im sure there would be some issues that you would need to work out.