Coding agents can use text-based skills to add domain knowledge and run repeatable workflows across repositories, browsers, tunnels, and messaging tools.
2
Chrome DevTools can expose browser actions and performance data to agents through MCP, while built-in AI features can explain errors and apply CSS changes back to source files.
3
Web AI APIs, LLMs.txt, and WebMCP let websites run AI locally, provide current documentation to agents, and expose application actions as callable tools.
Summary
Yohan Lasorsa and Olivier Leplus describe a web development workflow in which AI appears throughout the application lifecycle. Coding agents use repository skills and an agents.md file to inspect GitHub issues, run Playwright, create a tunnel, and send a preview to a phone. Chrome DevTools MCP lets an agent operate the browser, inspect console and network errors, run performance traces, and report issues such as oversized images. DevTools also has direct AI assistance for debugging and CSS changes, including applying edits to the workspace. The speakers then demonstrate experimental browser APIs for summarization, proofreading, and multimodal prompting with local models. These models can run in the browser without external API calls, although the APIs and browser support are still changing. The final section covers LLMs.txt for documentation discovery and WebMCP, which lets a site expose actions such as adding items to a cart or submitting a form as tools for agents.
Coding agents get better results when their workflows are described as reusable skills
Yohan Lasorsa says the main issue with coding agents in 2026 is often the user's skill in configuring them. Skills are lightweight text plugins based on an open specification supported by most coding agents. A skill can add domain knowledge, provide capabilities that are not built in, or describe a repeatable workflow. His Seine e-commerce demo has skills for GitHub CLI, front-end design, Playwright, a public tunnel, and sending Telegram messages. An agents.md file tells the agent to record a video after each change, run the development server, create a tunnel, send the URL, and leave the GitHub issue open until Yohan confirms the work.
Chrome DevTools MCP gives agents access to the browser's existing debugging tools
Olivier Leplus shows Chrome DevTools MCP as an MCP server that hosts browser tools an agent can call from an IDE or CLI. The available actions include clicking, filling forms, reading console messages, inspecting network requests, running Lighthouse audits, navigating pages, taking screenshots, and resizing the page. In the demo, an agent starts the application, opens Chrome, tests the main page, and runs performance checks under different network conditions. It produces metrics such as LCP and CLS, then points to problems including an oversized headphone image, CSS size, JavaScript issues, and JSON that could be preloaded.
AI built into DevTools can explain failures and edit source code from the debugging context
Yohan Lasorsa enables the AI features in Chrome DevTools and uses them on a CORS error in the console. The tool explains why the error occurs and suggests a fix without requiring a copy-paste into a separate chatbot. In the Network panel, an AI chat can inspect a failing request and explain that it is a bad request, possibly because an old endpoint no longer exists. In the Performance panel, AI analyzes a recorded GitHub trace and suggests ways to investigate render-blocking issues. Yohan also selects an H1 element, asks for a gradient matching the site's CSS variables, and applies the resulting live edit back to the workspace.
Browser Web AI APIs can run focused language tasks on the user's machine
Olivier Leplus introduces experimental Web AI APIs for summarizing, writing, rewriting, prompting, and proofreading. The summarizer can return formats such as a TLDR, teaser, key points, or headline, with controls for length and language. The model is downloaded to the computer and reused after the first download. In the demo, a review summary is generated locally in the browser. A proofreader corrects a review and returns both the corrected text and the changed spans with start and end indexes. The speakers stress that developers currently need Chrome flags, and the APIs may change as browser support develops.
Local browser models can accept images and return structured application data
Yohan Lasorsa uses a general language model API in the browser to generate a product review from an uploaded headphone image. The prompt asks for a one-sentence description of the product's condition, the user's reaction, and a title. A JSON schema constrains the response to a title and description. The resulting text describes a damaged headset and the user's disappointment, after which the review can be submitted. The model runs locally in the browser and accepts multimodal input such as text, images, and audio. The speakers present this as an alternative to sending requests to an external provider, while warning that the API is highly experimental and had changed recently.
LLM.txt gives agents a direct map to a site's current documentation
Yohan Lasorsa compares LLM.txt with existing files and systems that guide crawlers and site navigation. The proposal is a Markdown file containing links that help agents find the documentation they need without searching through every page. He shows Angular's LLM.txt as an example, with links to topics such as animation. The LLM-full.txt variant puts the site's contents into one large text file, including code examples. This can give a coding agent current framework guidance instead of relying on older information from model training. Yohan says the same pattern can apply to other types of content a website provides.
WebMCP lets a website expose business actions instead of making agents imitate clicks
Olivier Leplus explains that current browser agents often inspect screenshots or the DOM, guess which button to use, and click coordinates as a human would. WebMCP proposes exposing tools directly from the web application. His demo registers an add-to-cart tool with a name, description, input schema, and execute function. An extension or agent can call it with an item and quantity, and the page updates the cart. He also turns an existing review form into a tool, using its inputs and nearby labels to generate the schema. With auto-submit enabled, the tool fills and submits the form without additional human interaction. The specification is still highly experimental.
