So I installed mcp-cli into my AgentZero. This makes agentzero a basic smart mcp tool. Instead of context bloat and rot. It only injects the mcp-cli directions into the prompt roughly 400 tokens rather than upwards of 50k at times depending on how many mcps you have installed. I am now using mcp-cli only, i still have the original mcp framework, didnt need to overwrite anything. You can also set subagents to only be able to access certain mcps via mcp-cli. This has been a game changer for me because you can literally install virtually endless mcp servers without worry of context rot. Still doing some testing but working great as of now. When installing a new server I just ask the agent to install "xyz" to mcp-cli and then to test and verify all tools and functions in the new server to make sure everything works as it should. Try it out. Endless tools customizable for each individual agents via prompting the main agent. Also working on a improved RAG setup that works on top of the base memory and knowledge.

This is what I run in Claude Code. Goal is to have A0 manage as the overall Orchestrator of Orchestrators, if you will, via Terminals or subagents? We discussed a bit in weekly meet. Maybe this helps someone. Just takes small amount of customization to fit your needs! https://github.com/bybren-llc/wtfb-safe-agentic-workflow What This Is A production-tested Claude Code harness for teams that want structured AI workflows. Built on SAFe methodology (Scaled Agile Framework), adapted for AI agent teams. Works for any team with repeatable processes: Software, Marketing, Research, Legal, Operations. Includes: - 17 Model-Invoked Skills - Domain expertise Claude loads automatically - 23 Slash Commands - Workflow automation for common tasks - 11 SAFe Agent Profiles - Specialized roles with clear boundaries - Three-Layer Architecture - Hooks → Commands → Skills

Just finished my first A0 experiment and it's blown me away. Previously I did most of my copywriting using ChatGPT for ideation and Claude (via the web, then with the claude desktop) for the actual copywriting. Most of the time the copy was good, occassionally it was great (I've been writing copy for myself and for others for 15+years and have racked up multiple seven figures in client earnings as a result, so I'd like to think I know good copy when I read it), but at least once in every writing session it would make shit up, wildly hallucinate or just ignore some instructions. I took my best claude writing skill, moved it into A0 and started to write some linkedin pieces. Initially I noticed a few of the same errors but each time I asked for a correction the copy got tighter, the agent did seem to make less obvious mistakes. it took about 45 minutes of corrections, feedback and revisions but asking A0 to create and update an insights.md doc as it went appears to have sped up the learning phase. Really excited to see where how this develops next! And, in case it's not obvious, I am glad I am here :)