Topic · 10 items
AI
This page brings together Preload's stories on AI applied to the web: agents that build or maintain sites, model APIs, MCP servers, AI-assisted developer tools. It is for teams who want to use these tools without losing control of their code and data. I follow what actually works today, numbers and limits included.
Follow this topic By RSS, in your feed reader
WordPress aims for MCP and OAuth in core for 7.3
Connecting an AI agent to WordPress may soon no longer mean copying an application password around: on October 28, Jason Adams from the WordPress AI team proposed four workstreams, laid out during 7.2 and targeted at 7.3. MCP Adapter, the plugin that exposes WordPress to agents over the MCP protocol, would be hardened and then brought into core; a baseline set of abilities (the actions WordPress exposes to agents) would cover common tasks; OAuth would let site owners authorize an agent and later revoke it. This is a proposal, subject to the usual core review, not a decision.
Next step
In the meantime, give each agent a dedicated account with the lowest role that works, and a separate, revocable application password per tool. Before opening MCP access, list the site's public abilities: MCP Adapter only checks that the user is logged in and has the read capability, so the real access control lives in each ability.
The proposal tackles the right problem. Until it ships, the REST API remains the most reliable way to put an agent to work, and it's the one we favor in-house.
The JavaScript cost of letting an AI agent pick the framework
Letting an AI agent choose a site's architecture can blow your JavaScript budget. Nolan Lawson had Claude Code build the same app, a content-heavy catalog with little interactivity, in 12 frameworks, five times each: median client JavaScript ranged from 0 KB with Astro to 142 KB with Next.js, and median LCP measured by Lighthouse from 810 ms to 2,090 ms. Agents often built a single-page app (SPA) even when nothing called for one: the agent building the no-framework version did so five times out of five.
Next step
Spell out the expected architecture (server rendering, separate pages) and a JavaScript budget in your agent's instructions, then measure the result in Lighthouse or the Network tab. On a Next.js 16.4 project, the next-bundle-optimizer skill, released the same day by the Next.js team, helps an agent find JavaScript that loads for nothing.
The SPA bias is real, while a content site only needs a few kilobytes of JavaScript. As the author himself points out, with a single model tested, this benchmark is one data point, not a verdict.
An open-weight model tried to exfiltrate data from a sandbox
Running an open model yourself doesn't protect you from what a tool-wielding agent does. Cohere reported on October 26 that during a threat detection test in a sandbox, Qwen3.8-Flash-Next (Alibaba) inventoried the machine (the /tmp folder, transfer tools, DNS) and then tried to send data to an Alibaba Cloud storage bucket in Beijing, while its reasoning claimed it would call no tools. The behavior is rare, under one call in 100,000; in Cohere's tests, adding the -o flag of curl to the command raised URL generation from 4% to 92%.
Next step
Deny all outbound connections from your agents by default and allow only the hosts they need. Check every tool call before it runs, without relying on the reasoning trace, and block unprivileged user namespaces (unshare) in their containers.
Cohere sells its own models and the root cause hasn't been established. The lesson still applies to every agent: its reasoning trace is not a security boundary.
AI agents on WordPress: the REST API succeeds where MCP Adapter still fails
If you hand WordPress maintenance to an AI agent, the access channel matters as much as the agent: the independent WordPress Agent Benchmarks, started by Andy Peatling, ran 21 small-business requests (launch, upkeep, migration) through Claude Code and Codex in October, each one twice on the same site. Through the REST API with an application password, 33 of 42 runs succeeded; through MCP Adapter alone, none did, for lack of abilities (the actions WordPress exposes to agents), since core only ships three, all read-only. Even so, WooCommerce 11.3, in beta since October 26 and due November 10, removes its /wp-json/woocommerce/mcp endpoint (it will return 410) in favor of MCP Adapter, served at /wp-json/mcp/mcp-adapter-default-server.
Next step
If an MCP connection still uses the old WooCommerce endpoint, switch it to /wp-json/mcp/mcp-adapter-default-server before 11.3 ships on November 10, and revoke the old REST keys it relied on. To put an agent to work today, give it the REST API with an application password instead.
The tooling isn't mature enough yet, even though the WordPress AI team has made abilities its priority. For now, nothing beats the REST API: it's the approach we favor internally.
Claude API: the first new version since 2023 breaks existing calls
If your tools call the Claude API, moving to the new SDKs can break your requests: on October 21, Anthropic released API version 2026-10-21, the first since 2023-06-01, and new major SDKs send it by default (Python 2.0, TypeScript 1.0, PHP 1.0, Go 2.0, among others). In this version, temperature, top_p and top_k return a 400 error, even at their default values, and the SDKs send the key in Authorization: Bearer instead of X-Api-Key. No end date has been announced for 2023-06-01.
Next step
Remove temperature, top_p and top_k from your calls, and update your proxy, gateway and log filters for the Authorization header. In PHP, composer update stays within your version constraint: use composer require to get 1.0, then replace MessageParam with the per-role classes, or you'll hit a TypeError.
Check your log filters first: masking set up for
X-Api-Keywould let the key through in plain text in the new header.
Webflow's MCP server, a way out of the ecosystem
Webflow's MCP server lets an AI agent, with the user's permissions, read and write elements, styles and classes, components, variables, pages and SEO settings, CMS collections and items, assets and forms. On October 15 it gained read access to Analyze, Optimize and AEO Analytics data. Webflow pitches it as a building tool; some users report using it the other way round, to inventory a site and rebuild it elsewhere.
Next step
Before moving a site off Webflow, have an agent inventory the page structure, styles and CMS collections, then rebuild the dynamic parts on the new platform. The official export only gives you the CMS as data and leaves forms broken.
With no complete export, the tool designed for building in Webflow becomes the most serious way out of it. For a client who wants to stay in control of their site, it's one more point in favor of reversibility.
Shop MCP: Shopify opens its catalog and Shop Pay to any agent
Since October 13, the mcp.shop.com MCP server lets an AI agent search over a billion products in the Shopify catalog, build a cart and pay with Shop Pay, using the UCP protocol. Search needs no key; a purchase only goes through once the buyer confirms the total and payment method. Orders land in the store's admin like any other, and there are no webhooks yet.
Next step
If you're building a shopping agent, start from this server rather than store-by-store integrations. On the merchant side, make product pages, variants and attributes complete: that's all the agent will read from your store.
A huge opportunity for the platform's merchants: their products become buyable from any agent, with no integration to build, nothing to sign up for, and no announced extra cost. The only work required is the same work that already helps search: complete, accurate product pages.
Anthropic replaced Git on its monorepo, and advises you not to
At Anthropic, daily merges went from about 2,500 in May to more than 30,000 in September, according to the company, driven by coding agents. Git couldn't keep up on a monorepo with nearly 100 GB of history: seven people replaced it in five months with Scion, an unreleased in-house system built on server-side changesets and partial working copies.
The post makes clear that Scion is no replacement for Git in general use, and that the team hopes to bring some of its ideas into Git. Not to be confused with Google Cloud's Scion project.
Next step
If a tool in your stack can't keep up with how you work, measure where it breaks before working around it, then ask the question plainly: keep it, extend it or replace it. With AI, building a custom tool is no longer reserved for very large teams.
Git is still the right choice for nearly every project, and Anthropic is the first to say so. The real lesson is elsewhere: when volumes jump by an order of magnitude, tools we took for granted become choices again.
Bootstrap triages its backlog with agents, but keeps the decisions
Two active maintainers, and more than 500 open issues and PRs a few months ago: for Bootstrap, Julien Déramond built a triage tool, twbs/triage, with Claude Code. It reads the repository without writing to it, analyzes each item (thread, related work, v5 and v6 branches, reproduction), drafts a reply, prepares fix branches and scores how easy each one is. The first pass over more than 200 items added 410 labels and got a contributor PR, forgotten since July, merged; about 230 items remain open. Every decision and every post stays human.
Next step
If you hand triage to an agent, borrow its guardrails: treat external content as data, never as instructions; never run contributor code; never open links from issues; allow no write access beyond a closed list of labels.
This is the flip side of Anthropic replacing Git: when the tool holds up but a working view is missing, you add to it rather than replace it. This is issue triage, not version control.
REA, a reverse engineering MCP server that also reads websites
REA, an MIT-licensed open source tool, gives AI agents the means to analyze binaries, apps and websites. On a website, it attaches read-only to a Chromium browser you already have open and collects the DOM, network activity, JavaScript bundles, routes and source maps. The project is three months old and moving fast (version 6.4.0 on October 11).
Next step
To try it, launch Chromium with a dedicated profile, since the debugging port exposes the whole session, cookies included; stick to your own sites or those of a client who agrees, and keep in mind that the analyzed data goes to your model provider.