Rolling Out MCP Skill Loader Across All 35 MemStack Projects
Just wrapped up a massive deployment session that I've been putting off for weeks — getting the MCP Skill Loader properly configured across all my MemStack-linked projects.
The Challenge
I had 35 projects all linked to MemStack via NTFS junctions, but only StreamStack had the .mcp.json config file. Every other project was missing the MCP setup, which meant I wasn't getting the full benefit of the skill loader integration.
The trickiest part? Windows path handling. My initial approach using shell scripts for JSON generation was eating backslashes, which meant all the paths were broken. Classic Windows development pain point.
What I Built
Created a proper Node.js deployment script (deploy-mcp.js) that handles the JSON generation correctly and includes a --force flag for when I need to overwrite existing configs. Much cleaner than wrestling with shell escaping.
Also bumped start-memstack.bat from v3.2.1 to v4.0.0 — now it auto-deploys MCP configs whenever you run the link command. One less thing to remember.
The Rollout
Deployed .mcp.json files to 34 projects (StreamStack already had one) with proper Windows backslash paths. The good news is the trigger rule (memstack-mcp.md) was already in the shared .claude/rules/ directory via junction, so all projects were already getting that.
Tested on AdminStack, GreenAcres, and EpsteinScan — everything's working smoothly.
Design Decisions
Made the bat file smart enough to warn about existing MCP configs but not auto-merge them. If a project already has other MCP servers configured, it requires manual intervention. Better safe than sorry when you're dealing with 35 projects.
Also worth noting: MCP supplements the NTFS junction setup, doesn't replace it. The junctions are still doing the heavy lifting for shared rules and configs.
What's Next
Need to finish the StreamStack Railway deployment I started earlier, plus get Supabase set up since the jobs are currently just living in memory. Then I'll test the MCP Skill Loader connection across a few random projects to make sure everything's actually working in practice.
Feels good to have this infrastructure piece finally locked down across the entire ecosystem.