How FreeToolHub runs 21 AI tools entirely in the browser with no backend

·1 views

FreeToolHub has built and shipped 21 AI-powered tools — including email writers, resume rewriters, image upscalers, and audio transcription — without any server, API keys, or upload endpoints. All models download directly to the user's browser and run locally, meaning user data never leaves the device, a deliberate privacy-first product decision. The team uses Hugging Face's transformers.js on top of onnxruntime-web, with WebGPU as the primary backend and WASM as a fallback, and maintains a single TypeScript registry as the source of truth for all 18 models. A key technical challenge was enabling SharedArrayBuffer for threaded models without breaking third-party ad and analytics scripts, which was resolved by using the lesser-known Cross-Origin-Embedder-Policy: credentialless header instead of the standard require-corp approach. To manage large model sizes — some exceeding 1 GB — the team ships two quantizations per model, caches downloads in IndexedDB, and offers a smaller 200 MB sibling model for users on slower connections.

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)

How a self-inflicted cache poisoning made a live article return 404 for 23.8 hours

A development team discovered that one of their published articles was returning a 404 error exclusively to their own systems for nearly 24 hours, while all other users could access it normally. The root cause was a cache poisoning incident they had triggered themselves: their stats-collection job had requested the article's public URL before it was published, causing a CDN edge node to cache the 404 response. When the article went live, the origin server served it correctly to the world, but the team's runner kept receiving the stale cached 404 response — complete with headers showing a cache age of 85,673 seconds. A cache-busting query string failed to help because the edge node's cache key did not include query parameters, making it impossible for the team to bypass the stored response from their own network. A secondary bug in their error-reporting logic compounded the issue by making the single affected article appear as a 44-article blackout in their monitoring dashboard.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Import Word Docs into Confluence Cloud and Where the Tool Falls Short

Confluence Cloud has a built-in Word document import feature that requires no third-party app — users can access it via More Actions > Templates and Import on any page. The tool reliably converts headings, tables, and images, but drops shapes, text boxes, tracked changes, and headers or footers. Common import failures typically stem from poorly formatted source documents rather than Confluence itself, and can be reduced by applying proper heading styles, flattening drawn objects, and accepting tracked changes before importing. However, the built-in tool is limited to one file at a time and one file per page, making large-scale migrations slow and impractical. A folder of 150 documents, for instance, could realistically consume four full working days of manual effort, which is why many Confluence migrations stall before completion.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Lumiana lets developers use Node.js APIs directly alongside browser UI code

A developer has released an early-stage project called Lumiana that allows personal app UI and logic to be written together without building a separate API layer. The tool integrates with Vite and works by connecting the browser to a Node.js server using credentials, after which Node imports can be used alongside front-end code. JavaScript runs in the browser where possible, while OS-level operations are handled server-side. The project has been tested with libraries including Express, Sharp, node-pty, and sqlite3, though most packages do not yet function normally. The creator acknowledges the tool is still in early development, with significant work remaining before broader compatibility is achieved.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Key Questions Boards Must Answer Before Approving AI Coding Tool Deployments

Corporate boards are being urged to demand rigorous evidence before approving AI coding tool rollouts, going beyond vendor productivity claims. A July 2025 METR study found that experienced developers actually took 19% longer on tasks when using AI tools, despite believing they had worked faster. Experts recommend boards request baseline metrics covering defect rates, delivery times, and rework costs, alongside detailed maps of what data and systems the tools can access. Risk frameworks such as the NIST AI Risk Management Framework and Secure Software Development Framework are cited as guidance for structuring board-level oversight. Key warning signs include business cases built solely on vendor statistics and unverified claims that vendor systems do not train on company data.

0 comments Read more at DEV Community