
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Rachel Nabors’ core claim is that the browser is still an “infinite canvas” for agents — after years at Mozilla, W3C, Microsoft Edge, and React, she argues the same web primitives behind rich sites can power agent experiences too.
HTTP-hosted MCP servers make agent integration feel like adding a connector, not editing scary local JSON — she contrasts local “STDIO”/“studio” setups with web-hosted MCP endpoints that Claude users can add via a URL like /mcp.
Her comic archive became a practical demo for serving three audiences at once — one server supports humans in browsers, humans in agents, and agents in browsers, with tools like list comics, search by character, and a markdown-returning get transcripts.
MCP apps let a tool return a real UI, not just a wall of chat text — her get page tool renders an inline comic reader inside Claude, complete with comments, navigation, and a text mode transcript, using bundled HTML/CSS/JS in a sandboxed iframe.
The current MCP ecosystem has a big gap around resources and docs — she bluntly says agents should not be forced to call MCP tools just to shovel markdown docs into context, and asks client builders to finally expose MCP resources in the UI.
Web MCP is her bet on agents navigating pages without screenshot-guessing or DOM token burn — by exposing forms or JavaScript functions as tools through navigator.modelContext.registerTools, a browser agent can do things like “go to the next page” directly from the page itself.
Rachel Lee Nabors opens with a high-energy résumé lap: Mozilla, W3C, Microsoft Edge, React docs, and now Arize. Then she grounds it in a deeply human backstory — before big tech, she made web comics for teenage girls on iVillage, had 400,000 weekly readers, and learned to build for the web because that was how you reached people where they already were.
The talk’s project starts with procrastination: her 2010-era comic archive was breaking, with 404s and dead CDN assets, so she decided not just to repair it but to “future-proof” it. Her goal was delightfully ambitious and messy: make one server work for humans in browsers, humans in agents, and agents in browsers — “they’ve got like this polycule thing going on.”
She gives a practical distinction between local STDIO (“studio,” in her running joke) and HTTP transports for MCP. STDIO means command-line JSON and local process spawning that normal users hate; HTTP means a web service, POST requests, serverless friendliness, and a much simpler experience where a user can just paste a connector URL into Claude.
Her MCP server exposes tools that mirror the site itself: list comics, list storylines, list characters, search transcripts, search by character, and fetch transcripts in markdown. That symmetry is the point — don’t invent a weird parallel universe for agents; expose the structure your site already has.
This is where the talk gets spicy. Nabors says MCP resources are the right way to provide large text corpora like docs or comic transcripts, but the spec is loose and clients don’t surface them, so developers fall back to awful tool-based hacks that waste context and make models slurp markdown the hard way.
get page: the moment the agent stops being just a chat boxHer favorite demo is get page, a tool that returns an MCP app — a mini comic reader inside Claude with visuals, comments, next/previous navigation, and a text-mode transcript toggle. It’s her argument in one artifact: chat is the “lowest common denominator,” and users shouldn’t have to discover everything by interrogating a blank starfish-shaped text box.
She walks through the implementation details with equal parts excitement and warning. MCP apps are single HTML files in sandboxed iframes, so assets often need bundling or base64 embedding, external resources need CSP/CORS care, there’s no localStorage, no direct network access, and links require host permission — a recurring “mother may I?” pattern.
She closes on Web MCP, which exposes forms or JS functions to in-browser agents so they can act on a page without screenshotting it or chewing through DOM/XML tokens. Using navigator.modelContext.registerTools, she demos a browser agent reading a comic transcript and moving to the next page, then lands the bigger thesis: the browser isn’t dying, it’s evolving, and APIs like Web Speech, Audio, Canvas, WASM, and CSS are all still there waiting to be used.
Share
Keep Reading
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.

Playbook
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.

Playbook
OpenAI shipped /goal for the Codex CLI. It turns a prompt into a persisted, self-continuing contract.