Developer builds visible, local browser for AI coding agents to test sites

·1 views

A developer needed to manually test a platform under varied conditions, such as slow networks, different screen sizes, and signed-in flows, and wanted an AI agent to handle it. Existing tools fell short: Playwright MCP's browser closes with the agent session and runs unwatched, Chrome extensions expose a personal browser, and cloud browsers cannot reach localhost or internal staging. The developer therefore built Hronaut, a visible Electron desktop browser that an AI agent drives through a local MCP server. The browser keeps running when the agent disconnects, lets a human pause or take over, isolates each task in its own profile, and filters passwords and secrets from what the agent reads. It runs on Windows, macOS, and Linux and works with any MCP-compatible agent.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Shopify 2026-10 API Deprecates owner_type, Requires Global IDs in metafieldsSet

A developer warns that the Shopify 2026-10 API version contains a breaking change to the metafieldsSet mutation that can cause failures in production pipelines once the deprecation window closes. The owner_type argument, which previously let developers specify resource kinds for batch metafield writes, is deprecated in this version. Shopify now requires the ownerId field to be a fully qualified Global ID that matches the metafield definition's namespace. Payloads that fail this check may return an INVALID_OWNER error or be silently dropped, depending on how an integration handles errors. The author recommends constructing Global IDs explicitly in client code and treating such errors as fatal rather than as warnings.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo developer details lessons from building a scheduled short-video publishing pipeline

A solo developer spent eight weeks building PilotReels, a tool that generates daily faceless short videos and posts them to YouTube Shorts, Instagram Reels, and TikTok. After an early bug caused duplicate hooks to publish, the tool added a manual approval step where users can edit scripts before rendering, which users said built trust. A second AI model judges generated scripts against a rubric and triggers regeneration when they fall short. Captions are built from timestamp-aligned voice output, and Arabic text required fixes for bidirectional ordering, line wrapping, and punctuation placement. Each pipeline stage stores its output under an input-derived key so retries skip completed work, and the system runs on a deliberately simple stack with a Postgres-backed job queue on a single cloud instance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

LazyNotes AI Offers Open-Source Study Tools to Cut Exam Prep Time

A developer has built LazyNotes AI, an open-source study helper designed to reduce the time students spend preparing for exams. The app condenses notes into summaries, extracts key points and terms, generates revision questions, and offers beginner-friendly explanations of difficult concepts. It is built with Python, Streamlit for the interface, and Ollama for AI inference, supporting both local models and a hosted API for a public web version. The developer highlights open-weight models as a way to offer flexibility and control over where notes are processed, though the hosted version requires an internet connection. The project was created for the Hacktoberfest Open-Source AI Challenge, and its time-saving benefits have not yet been measured, with testing against real study notes planned next.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Research Answers With Sources Can Still Be Wrong

A developer building the research engine for the Prizston platform argues that an AI answer can include citations and still reach a conclusion the evidence does not support. He describes questions researchers must ask, such as when sources were originally published and whether they actually back up a claim. He uses a constructed example in which a page lists two conflicting publication dates, and says the correct result in that case is "not_determined" rather than picking the newer date. He proposes separating source observation, evidence evaluation, and answer validation as distinct stages. He also suggests testing whether a conclusion holds up after removing its strongest supporting source.

0 comments Read more at DEV Community