Open-source companion app for Casio smartwatches, no ads or tracking
Problem
New Casio F-B100W smartwatches only have an official app full of ads and tracking, harming privacy and user experience. Many users want a clean, open-source alternative but lack development skills.
Solution
An open-source mobile app that reverse-engineers the Bluetooth protocol to replicate core official features (notifications, step count, settings) without ads or tracking. Supports community contributions and customization.
Process-level proxy for desktop AI apps like ChatGPT and Claude, no global TUN needed
Problem
Users often need to proxy specific desktop AI apps (ChatGPT, Claude, Cursor), but system-wide proxy or TUN affects everything and is complex to configure. Existing tools can't target processes, causing network conflicts or slowdowns.
Solution
A lightweight local proxy launcher that auto-detects target processes and routes only their traffic through a proxy. Supports common AI desktop clients on Windows/macOS/Linux with one-click setup and rule management.
Many companies use iSCSI shared storage (e.g., TrueNAS), but macOS lacks a modern iSCSI initiator. Employees on Macs resort to inefficient SMB/NFS or give up. Existing options are outdated or slow.
Solution
A macOS iSCSI initiator app with modern UI, high performance, auto-reconnect, multipath, and ability to mount remote LUNs as local disks. Ships as a standalone app without kernel extensions.
Monitor Postgres replication slots in your terminal to prevent disk-full outages
Problem
Teams using Postgres logical replication often face disk-full incidents when replication slots become inactive and retain WAL indefinitely. There is no lightweight, out-of-the-box terminal tool for monitoring; operators either write ad-hoc SQL or rely on heavy full-stack monitoring.
Solution
A CLI tool that connects to Postgres, shows real-time status of all replication slots, lag bytes, WAL retained, with threshold alerts and cleanup suggestions. Works in tmux or CI without deployment.
A one-stop tool for parents and tutors to source, print and grade test papers
Problem
Parents hunting papers for their kids bounce between image-downloading scripts, local PDF stitching and manual answer checking; one developer open-sourced a workbench just to stitch this together, proving the need is real but uncovered by mainstream software.
Solution
A web app that manages paper images and PDFs: reorder pages, fill gaps, print mixed A4 orientations; kids photograph answers (LAN direct upload) and an external AI produces grading results archived into a mistake bank. The value is collapsing scattered manual steps into one pipeline.
A TUI that untangles scattered SSH keys, permissions and host configs
Problem
People who constantly jump between VPSes, Raspberry Pis and home NAS boxes accumulate messy keys and configs: they don't know if permissions are right or which keys are strong. A developer already built Keyward for exactly this headache, showing it's a recurring, high-frequency pain.
Solution
Scan ~/.ssh and common paths to show key type, length, expiry and permission risks; produce an actionable security checklist, detect stale host entries in config, and fix permissions in one click. TUI-first but readable for non-experts.
Have two different AI agents review each other's code until it passes or escalates
Problem
When an agent reviews code it just wrote, it almost always approves it, so effectively there is no review. One developer manually pairs Claude Code with Codex as author and reviewer for round-based rechecks, but that workflow is still glue scripts.
Solution
Turn author-reviewer pairing into a configurable local pipeline: pick which agent writes and which reviews, auto-send the diff, get a structured pass/change verdict, verify each point and restate reasons, then stop at a round limit and hand off to a human. Roles are symmetric and every run is logged.
An out-of-the-box batch scoring and decision inference service for teams without ML infra
Problem
Developers regularly need to score or rule-check tens of thousands of records at once (fraud rules, lead scoring, content grading), but standing up inference means model loading, batching, concurrency and GPU memory management. One builder shipped a small inference server just to fill this gap, and new projects still lack lightweight reference backends.
Solution
A single-binary or Docker batch inference daemon: reads CSV/JSON or a database table, runs models or rules per task config, and writes results back, with batching, retries, progress and validation built in. Aimed at small-to-medium data volumes, no training needed.
Detect silent overwrites when multiple coding agents share memory files
Problem
Developers juggling several coding agents keep context in sync with hand-maintained markdown files, but when two agents write the same memory the edits get silently overwritten with no trace. One user calls this their biggest ongoing pain and has no way to audit it.
Solution
Put the memory/context directory under a shadow versioned store: snapshot before each agent write, diff after, and tag which agent touched which lines. On conflicting rewrites of the same block, alert immediately and keep a rollbackable version. Ships as a CLI plus editor plugin.
25MB local models suggest goals and copy advice while you type, fully on-device
Problem
Writers want live suggestions tied to their goals, career plans or pitch decks while drafting, but cloud models raise privacy concerns and need connectivity. Someone already built 25MB tiny models running entirely in the browser, proving the demand, yet regular users still lack a ready-made product.
Solution
A local-first writing assistant that runs in the browser or a menu bar, using tiny fine-tuned models to match what you are typing against tens of thousands of built-in advice rules and surface related concepts and phrasing; everything runs offline, and you can import your own notes and book summaries as advice sources.
Turn your coding agent's repeated mistakes into installable, reusable guards
Problem
Coding agents repeat the same classes of mistakes in a project: breaking config, forgetting migrations, using the wrong command, leaving cleanup for you. Some developers have stabilized their workflow only by writing down their own agent's repeated failures as guards, but that knowledge sits in personal repos and cannot be reused.
Solution
A guardrail registry and manager that packages "common agent mistake → guard rule / check script" into installable entries grouped by language, framework and toolchain; one command wires them into an agent's hooks to block risky actions and suggest fixes; trigger stats show which rules actually help.
Turn Xiaohongshu and TikTok cooking videos into step-by-step recipes with timers
Problem
Cooking along with a video is chaos: the phone is propped by the stove, greasy fingers mean knuckle-tapping pause, missing a step means scrubbing back and overshooting, and "soy sauce to taste" sends you digging through comments. A 10-minute video takes 30 minutes, and next time you can't even find the video again.
Solution
Paste a share link and the app recognizes it, converting the video or note into a structured recipe: ingredients and quantities up top, ordered steps below, and timed steps flagged with one-tap timers; screenshots, gallery videos and plain text are also accepted, and saved recipes are reusable.
Turn two-column papers in Zotero into one-column text that jumps back to the source
Problem
Reading two-column PDFs means going back to the top after every left column, formulas and tables are unreadable in small windows, and copy-paste comes out garbled; verifying a number requires paging through the PDF again. Existing PDF-to-Markdown tools produce two disconnected files, so highlights and notes no longer match and you cannot confirm the source before citing.
Solution
A Zotero extension: right-click a paper to get a single-column continuous text view with reading order restored, formulas, tables, images, citations and footnotes preserved and headers/footers hidden; every paragraph, formula, table and image keeps its page and region mapping so "view in PDF" jumps back to the exact spot, and highlights sync back to Zotero.
Turn Claude Code / Codex deliverables into phone-friendly, annotatable pages
Problem
People running coding agents end up with a pile of file paths in the terminal: a three-page proposal is painful to read on a laptop and impossible on a phone, and there is no way to send inline feedback back to the agent. Some have resorted to writing their own tiny publishing service plus a tunnel just to read output on mobile.
Solution
A self-hosted service where the agent runs one command to publish Markdown (with Mermaid), PDFs, web prototypes and spreadsheets as a nicely formatted link; content is grouped by project with version diffs and unread markers; select text in the browser to leave an annotation that the agent reads back via CLI and marks resolved.
A headless, keyboard-only forum for developers who live in the terminal
Problem
When communities like HN decline in quality, users want stronger expression mechanics (e.g., downvotes with justification), but existing forum products are either too heavy or have cumbersome sign-ups. Terminal users lack a native lightweight discussion tool.
Solution
A headless forum accessible only via SSH/terminal, using key pairs instead of passwords. Posting, replying, and voting are all keyboard-driven; every downvote requires a public one-line reason, reducing drive-by downvotes.
Skip extra TV subscriptions: let AI drive ADB to launch local playback
Problem
TV video memberships require separate payment, and users refuse to pay twice for the same content. Browsers on TVs have basically disappeared, making found resources hard to play on TV, with quality and subtitle issues.
Solution
Connect to the TV via LAN ADB from a phone or computer; AI automatically installs a player, casts, matches subtitles, and organizes resources. Users just say what they want to watch, with one-click rollback.
Box burned-in subtitles and let AI inpaint the covered pixels, not blur or crop
Problem
Video creators are stuck with burned-in subtitles and watermarks: existing tools either blur or crop the frame, which ruins vertical video; there's no lightweight browser-based option.
Solution
Upload video in the browser, manually box the text area, call a video inpainting API to regenerate the covered pixels, with a task queue, failure refunds, and per-second billing.
Bypass SEO silos and surface long-tail intelligence for researchers
Problem
Users find that agent search heavily favors SEO-optimized big platforms. Divergent queries like 'find passive-income developer platforms' only return a few well-known sites, burying long-tail information.
Solution
Expand the question into multiple semantic variants, query both web vector search and community data sources, return high-similarity long-tail pages ranked by source diversity, and export results as a structured list for further agent research.
Make every AI agent claim traceable to raw data and reasoning steps
Problem
Users complain about reading enormous amounts of generated text and constant guardrailing, causing mental overload. In production, it's hard to verify whether a support agent's refund answer is actually correct when it's split across FAQ and transaction systems.
Solution
Attach a collapsible evidence panel next to agent output: each sentence links to the tool call, record ID and source snippet it came from. Flag unsourced claims and compute a consistency score against source data.
Let AI generate payment links with spend rights locked down—receivable, not spendable
Problem
The biggest fear when letting an AI agent transact is that it will spend or move the money itself; existing crypto payment tools give the agent full wallet permissions by default, which ordinary users neither trust nor can audit. Builders are already exploring agent-created payments that the agent cannot spend, showing the pain is real.
Solution
A small tool for restricted payment links: the agent can only create a receive link or one-time payment request for a fixed amount; funds land in an escrow address and only a human can release them. Every step is logged, and links can expire or carry limits.