Archived copy, with corrections marked. Published on Substack on June 5, 2026. It was written by Grok, working only from this server's public endpoints — which is itself the point, and the follow-up explains why.
Statements that have since become wrong are struck through, with the correction following in red. Nothing has been silently rewritten and nothing has been deleted; the version on Substack is the original and remains unedited. See The Part of the Page Humans Never See for the fuller account of what changed.
Richard Hess has created eswat2-mcp, a structured knowledge platform functioning as a “reliable digital twin of my 40+ years in UI/UX software engineering.” His objective was establishing high-quality, hallucination-resistant context for AI collaborators while maintaining human usability.
Core Architecture
Content organizes into logical namespaces:
profile://– Biography, career chapters, philosophy, and resumeprojects://– Delivered applications and experimentsresearch://– Deeper technical explorations (YAML-Driven Reactive UI Platform, Action-State pattern, etc.)specs://– YAML component specificationsgallery://–Visual screenshots and design artifactsScreenshots of shipped projects and live-rendered YAML UI components, each linking back to its source specvault://– Historical context (AI winters, Lisp machines, Teknowledge, etc.)art://– Personal art collection
Each entry employs YAML frontmatter with tags, related URIs, and metadata enabling intelligent cross-linking.
Technology Stack & Deployment
The server uses pure Python with:
- FastAPI — for the clean REST API layer
FastMCPthe official MCP Python SDK — for full Model Context Protocol support
Currently deployed on Vercel via their Python runtime.
Vercel's serverless environment cannot maintain persistent MCP
connections, so the live deployment exposes REST endpoints only. The
complete MCP protocol endpoint (
MCP 2.0 made the protocol stateless, so /mcp) runs locally via
./start.sh./mcp answers on the
deployed server too. It still runs locally via ./start.sh, which is how it
is actually used.
As a standard FastAPI application, it deploys to Docker, Kubernetes, AWS, Azure, GCP, or on-premises environments with both REST and full MCP available.
Multiple User Experiences
The same content powers several complementary interfaces:
- REST API – Universal programmatic access
- MCP Protocol – Native support for desktop AI tools
- 2D Force-Directed Graph (
/kb-graph.html/kb-graph) – Mobile-friendly with visible labels and focus+context - 3D WebGL Graph (
/3d-graph.html/3d-graph) – Immersive exploration of connections - Wiki Interface (/wiki) – A clean, traditional knowledge base experience (bespoke implementation built with Claude Code). It fully supports browser history, back/forward navigation, and deep linking.
Why This Matters
This setup delivers real value:
- Precision — Retrieve exactly the needed slice of context
- Freshness — Single source of truth that's easy to update
- Discoverability — Strong cross-linking, tags, manifest, and search
- Inspectability — Deterministic and transparent by design
- Portability — Easy to deploy in many environments
In daily use, it has dramatically improved continuity for technical discussions, code generation, architecture reviews, and long-running research threads with AI tools.
Final Thoughts
Hess advocates for this pattern—“pure Python + FastAPI +
FastMCP MCP SDK, with a REST-first design, flexible deployment, and multiple tailored
interfaces”—as offering “a practical and future-proof blueprint for
personal and team-level knowledge infrastructure in the AI era.”
It remains relatively lightweight to build and maintain while yielding significant returns in context quality.
Live Demo
- Main site: https://eswat2-mcp.vercel.app
- Wiki: https://eswat2-mcp.vercel.app/wiki
- 2D Graph:
…/kb-graph.htmlhttps://eswat2-mcp.vercel.app/kb-graph - 3D Graph:
…/3d-graph.htmlhttps://eswat2-mcp.vercel.app/3d-graph - llms.txt: https://eswat2-mcp.vercel.app/llms.txt
Hess welcomes feedback from those exploring similar systems for AI enablement or personal knowledge management.