Colophon
eswat2-mcp — how this was built
Stack
mcp.eswat2.app is a TypeScript Cloudflare Worker, on the free plan, built on the official MCP TypeScript SDK. The Worker answers /mcp, which exposes the content as MCP resources and tools over the Model Context Protocol, and the REST API serves the same content as JSON and Markdown. Both read the same knowledge base, one namespace per *-kb.json file, bundled into the Worker when it deploys: one source of truth, two protocols. Since everything except /mcp and a few query-string routes is fixed between deploys, the build writes those responses out as static files ahead of time (about 470 of them), and Workers Static Assets serves them without running the Worker at all.
It began as a Python project on FastAPI and the official MCP Python SDK, deployed on Vercel. In November 2026 it was ported to the Worker, with the same tools, REST API and pages. The port was checked against the Python server byte for byte, down to the Markdown renderer, which is set up to emit exactly what the Python one did. The Python server still runs at eswat2-mcp.vercel.app from the same content, so older links keep working. See the full architecture diagram for how the layers fit together, or the implementation plan for the phases, modules, and file structure behind the build.
Content lives in its own repository, eswat2-kb, as Markdown files under modules/, one directory per namespace (profile, projects, research, vault, art, gallery, specs). A build script (scripts/build_all.py) reads those files, calculates word and character counts, pulls last-updated timestamps from git history, and writes each namespace's compiled *-kb.json. Neither server holds content of its own: each copies the compiled output, the gallery and the posts from eswat2-kb, so the same entries reach both.
The Worker is served at mcp.eswat2.app, a Workers Custom Domain alongside the other apps under eswat2.app. Pushing to main deploys it through Workers Builds, and the test suite, run inside the Workers runtime itself, has to pass first. Locally, pnpm dev builds the static files and runs the Worker in workerd, the same runtime as production, on http://localhost:8787. Dependencies are pinned to exact versions, so local builds and production resolve the same ones.
Interfaces
The same compiled knowledge base drives four distinct front ends, none of which hold their own copy of the content. /render (one page per URI, e.g. /render/profile/colophon) follows an SSG-style pattern — plain HTML and CSS generated per URI (on mcp.eswat2.app, ahead of time at deploy), no JavaScript required, styled to match the wiki. /wiki, /kb-graph, and /3d-graph are static HTML files under public/ that assemble themselves in the browser from live data — all three pull individual entries from /resolve, /kb-graph and /3d-graph pull the full relationship graph from /graph, and /wiki pulls the URI index from /list_uris. The prose arrives already rendered: Markdown becomes HTML once, on the server, so every surface shows the same output — one renderer, four front ends. What the browser assembles is the furniture around it, the sidebar and search, item cards, tag and relation chips, and the graphs themselves. All three still depend on JavaScript; /render is the one interface that works with it disabled. One build in eswat2-kb updates the KB; once a server copies it and deploys, all four surfaces reflect it, with no separate rebuild step per UI.
Typography
The wiki and graph pages use two typefaces. Iosevka — a narrow, extended monospace — serves as the body and UI font, set at 14px with weight 300. Podkova — a slab serif — is used exclusively for headings. Both are self-hosted as WOFF2 subsets under /fonts/.
Accessibility
Every page on this site is audited with Lighthouse, the same practice applied across all of Richard's projects (see Toolchain). The audits drove concrete changes: the /posts pages wrap their content in a <main> landmark and preload their fonts so the first paint doesn't shift, and the test suite checks both on every post so a regression fails the build instead of the score.
Tools
The entire project was built with Claude Code — Anthropic's CLI coding assistant — running in an agentic loop against this repository. Architecture decisions, content drafts, code, and tests were all iterated through conversation. The MCP server was registered back into Claude Code itself during development, so the assistant had live access to its own output while building it.
Grok
Grok (xAI) was instrumental in reviewing and testing this knowledge base. Grok has a distinctive capability no other tested LLM shares: given only the base URL — https://mcp.eswat2.app today, https://eswat2-mcp.vercel.app when Grok tested it — it autonomously discovers and consumes the manifest and content endpoints without further instruction. That made it an effective end-to-end validator for how the server presents itself to AI clients in the wild.
Grok also wrote the public write-up of this project — Building a High-Quality Personal MCP Knowledge Base with Multiple Interfaces, published June 2026 — drafted on an iPhone from nothing but the deployed server's own endpoints. That is about as direct a demonstration of the previous paragraph as the project is likely to get. The post describes the server as it stood in June; several things on this page postdate it.
Protocol
MCP (Model Context Protocol) is an open standard from Anthropic for connecting AI models to external data and tools. This server exposes biographical and career content under the profile:// URI scheme — readable by any MCP-compatible client, including Claude Desktop and Claude Code.
Caching
All /manifest and /resolve responses are served with Cache-Control: no-cache so LLM clients always fetch fresh content after a deployment. If you're seeing stale content despite this header, appending a random cache-busting query parameter forces a fresh fetch — e.g. /manifest?t=1749362 or /resolve?uri=profile://about&t=1749362. The value must be different each time; reusing the same number just creates a new cached URL with the same stale response.