Writing: MCP Apps & Agentic Software

2026-07-16 A source-aware AI writing workflow: citations are not enough How to separate discovery from authority, map claims to evidence, and stop unsupported drafts before publication.

2026-07-16 WebMCP vs. computer use: what the OLX test actually proved One checkout, two paths, and the exact point where interface interpretation became the bottleneck.

2026-07-15 ChatGPT Apps became plugins: what actually changed The app is still the MCP-backed capability; the plugin is the crate it ships in. What changed, what to migrate, and what not to rewrite.

2026-07-15 ChatGPT Visualize vs. Claude Artifacts vs. Sites Visualize answers, Artifacts persist, Sites deploy: how to pick between three products that all turn a prompt into something clickable.

2026-07-15 Claude Live Artifacts vs. MCP Apps A host-generated view for one person, or a service-shipped interface for every host. The whole comparison is one inversion.

2026-07-13 WebMCP vs. MCP Apps: two paths to agent-enabled software WebMCP makes the agent a guest in your website; an MCP App makes your interface a guest in the agent host. Two standards, two surfaces.

2026-07-13 WebMCP vs. APIs, MCP, and browser automation Five rungs separate a direct API from a model reading pixels. Three WebMCP Teach experiments show why the rungs are not interchangeable.

2026-07-13 A production case study: human-approved Google Sheets diffs The unit of trust in a spreadsheet agent is the patch: exact range, before, after, one human click, and a server that checks for drift.

2026-07-13 How to make MCP widgets work across ChatGPT and Claude A widget that renders in ChatGPT and goes blank in Claude usually broke one of five strings. The debugging order for finding which.

2026-07-13 ChatGPT Apps SDK vs. Claude MCP Apps Both hosts render the same MCP Apps standard. Build one server, one widget, one approval flow, and feature-detect the rest.

2026-07-13 How to build an MCP app with human approval OAuth grants account access, not approval for a specific write. The four-layer architecture behind exact-effect review and server enforcement.

2026-07-04 Building this portfolio with three agents The colophon: one odd reference, design decided in Paper, Claude Code holding the shape, Codex doing the DNS from a browser tab.

WRITING