Pointcast

Talk and point at your web app: your coding agent gets a spec with the exact elements and the code behind them.

Talk about the UI changes you want while you Alt+click the parts of your web app you mean. Pointcast turns it into a Markdown spec your coding agent can act on: what you said, the exact elements you pointed at, and the code behind each one. One recording can hold a whole list of changes. HOW IT WORKS 1. Press Record and talk, the way you would to a colleague: "this should take you to the reports page", "this button should export the order status". Prefer typing? Switch to Typed mode and write a short note next to each element instead. 2. Alt+click (Option+click on macOS) or select whatever you are talking about as you say it. Pointing never triggers the page: Alt+click on "Delete" never deletes. 3. Press Stop. A few seconds later the spec is ready: one request per sentence, each with the elements you pointed at while saying it. YOUR AGENT PICKS IT UP With the companion plugins for Claude Code, Codex and Gemini CLI (or Pointcast's MCP server in any MCP client), type /pointcast watch once. From then on, each recording goes straight to your agent when you press Stop: it starts working, tells you what it changed, and waits for the next one. Or type /pointcast to apply your latest recording. No integration? The spec is also on your clipboard: paste it into Cursor, Claude Code or any coding agent. IT BUILDS WHAT YOU MEAN The spec tells your agent that the elements you pointed at say where. A request about an existing element (its text, size or colour) changes exactly that; a request for something new (a behaviour, a component, content) is built properly, in your app's own style. Prefer strict edits? One setting, "Change only what I point at", keeps the agent to the elements you pointed at, and only as asked. IT POINTS AT THE CODE On development builds of React (on Vite, including React 19, and on Next.js), Vue 3 and Svelte 5, and on Django templates with the companion pointcast-django package, each element leads with its code: where that instance is used, and the line of its text or data in your source, quoted. Your agent goes straight to the right line instead of searching the codebase. On Next.js, the text and data lines come from your repository through the agent integration. - Point inside charts and maps drawn in SVG: one bar of a chart, one point on a map. - Point at several copies of one component (the rows of a list, a set of cards) and they are grouped into one entry instead of repeated. - The errors around the moment you pointed (a failed request, a console error) come with the element, so the agent knows what actually broke. One setting turns this off. PRIVATE BY DEFAULT - Your voice is transcribed on your device: Whisper runs inside the extension. No account, no server, no API key. Choose Fast (the default, 294 MB download) or Accurate (whisper-small, 512 MB: fewer misheard words, about twice as slow). - Works on local development hosts (localhost, 127.0.0.1, *.localhost, *.test) out of the box. Any other site needs your explicit opt-in, one site at a time. - Password fields and other sensitive inputs are never captured. On non-local sites, text that looks like personal data (emails, phone numbers, tokens) is redacted. - Open source (MIT): https://github.com/Hugelidus/pointcast WHY POINTING HELPS In an evaluation on three real open-source dashboards (React, Vue, Svelte), pointing raised the share of requests where the agent found the right element from 78% to 89%, exactly where words alone are ambiguous: two "Export" buttons, identical cards, shared components. A follow-up measured a recording of six changes against the same six changes typed by hand: 96% accuracy against 85%, with 24% fewer input tokens and 75% fewer searches. A third evaluation gave the same ten changes to the same agent on a real React app: once as one Pointcast voice recording (a text-to-speech voice, with the pointing automated so it can be repeated), and once as a carefully hand-written prompt. The recording took 1.5 minutes; the prompt took an estimated 15 to 20 minutes to write. In a blind review, the result from the recording scored higher: 192 of 200 against 174. That is one run of each, reviewed by an AI model, so read it as a direction rather than an exact margin. Methods, numbers and limits of all three are in the repository. Works in Chrome and Microsoft Edge on desktop; tested in CI on Windows, macOS and Linux. This is a beta: feedback and bug reports are welcome at https://github.com/Hugelidus/pointcast/issues The first recording downloads the speech model once (about 294 MB; 512 MB more if you switch to Accurate).