Thinking about a quick detour on the WOVEN x15 beta to take a serious look at MCP and some new mounting options.
If you’re just jumping in: WOVEN is an agentic system I’ve been building since December. The core idea came out of my own creative workflow—I wanted agents that could self-improve over time and learn from experience.
There’s been some debate about the alignment risks there, but my feeling is that it's a developmental and ergonomic problem, not a top-down control issue.
From memory to tooling
x13 was the first functional build, with memory fully wired up—but it was brittle and too monolithic for the task. It still works, though.
x14 added tooling and completely rethought the memory system. The new design uses a vector database and a chart graph ranking system to prioritize memories. I may have borrowed a few ideas from how Google’s PageRank works—but here, without links, it’s applied a little differently.
Getting tooling right was its own challenge—especially around long-running agentic loops, where you’re trying to support multi-step tool use without falling into the recursive repetition traps that plague a lot of other systems.
x15 introduced a dynamic configuration layer to simplify API integrations.
It was working great—until, just as I wrapped it... MCP dropped.
The MCP tradeoff
At first I resisted. MCP didn’t feel like good engineering to me. It felt sloppy.
But the more I looked, the more I saw the potential.
Right now, x15 uses OpenAI’s tooling framework. The advantage there, is that it's widely supported and available everywhere.
I even came up with a nice easy YAML-based system for building integrations—but you still need to write one per endpoint. It’s miles ahead of hand-wiring everything manually... but nowhere near as seamless as MCP.
The tradeoff?
MCP requires always-on persistent connections to every API it manages. That introduces real complexity around dropped connections, service restarts, CI/CD pipelines, and so on. But it radically simplifies streaming and dynamic app design.
You’re basically building an API to manage your APIs with an API—but if it works with zero setup? That’s a win.
This is going to throw off my totally arbitrary self-imposed timeline.
I was planning to launch a closed beta next week, followed by feedback and an open source release.
If I pivot to MCP and build out x16, with this alternate tooling layer and five new service integrations, I’m looking at another three-week delay.
Let’s just hope nobody radically redefines AI tooling again before I get it out.
Fingers crossed.