Hidden gems, rated before they trend.
2371 new projects found this period, 10 made the list.
Window 2026-09-14 – 2026-09-20Rated 1446
It converts real locations into playable maps for Project Zomboid B42 and explicitly credits upstream Knoxify; a rare genuine need for PZ modders and players who want to fight zombies outside their home, limited by dependence on the upstream toolchain and output quality that varies with geographic data.
KnoxMap: turn any real place into a playable Project Zomboid Build 42 map (built on Knoxify)
★34
It hash-chains AI agent tool-call records, adds deterministic replay and out-of-scope detection, filling the forensics and compliance audit gap after agents go live. 171 stars in 0.6 days show agent observability is missing exactly this piece. The zero-dependency CLI is easy for CI and regulatory workflows to adopt, with enterprise audit subscriptions as the path; the risk is trace platforms like LangSmith covering it incidentally.
Audit AI agent tool-call transcripts: hash-chain sealing, deterministic replay, and scope overreach checks. Dependency-free Python CLI.
★171
Bespoke Labs extends its synthetic data pipeline into a combination of local typed decisions, contrastive data filtering and model evaluation, hitting the real pain of fine-tuning data quality; 421 stars in a day and a half plus an existing Curator user base support a hosted data and evaluation platform, with the risk that big players like Databricks push similar capabilities down-market.
Local typed decisions, contrastive data curation, and model evaluation.
★421
Wraps any open-source model into a classifier or jev endpoint, capturing migration demand for the new typed decision paradigm, with 282 stars in a day and a half showing a clear gap in the ecosystem; it could charge for a gateway service or SDK, but the risk is that once Jev officially supports this natively, such thin wrappers get absorbed upstream.
Turn any open model into a classifier/jev endpoint
★282
Runtime behavior ablation strips refusals from a ternary-quantized 27B model without changing weights or requantizing, addressing how compressed models' safety alignment gets baked into weights and becomes hard to adjust. 360 stars in 0.7 days came from the uncensored topic. It could monetize through OrcaRouter inference routing, but risks compliance and model license disputes, and many similar de-alignment tools already exist.
Runtime behavioral ablation for compressed LLMs. First target: Ternary Bonsai 2 27B — no weight modification or re-quantization. by OrcaRouter team
★360
Detects repeated segments in first-person video and locates their time ranges, real dirty work in robot data cleaning, with a narrow scope and accurate description. The limit is it only helps egocentric dataset builders, and you'd need to confirm an installable package was actually released.
Local-first detection and temporal localization of reused footage in egocentric video.
★33
Brings ANGLE's Metal/D3D11/Vulkan backends into Node, enabling offscreen WebGL and three.js rendering without a browser, usable today for server-side screenshots and visualization output. It adds WebGL2 and multiple backends over headless-gl, at the cost of needing your own native build environment.
Headless WebGL 1 & 2 for Node.js, powered by ANGLE (Metal, D3D11, Vulkan). Render any WebGL or three.js scene without a browser.
★32
Flashes a second-gen Echo Dot with fully local firmware, removing Alexa, enabling all seven microphones, and connecting to Home Assistant as a voice satellite, with signed A/B updates and SSH. Hardware tinkerers will want to try it immediately, but it requires disassembly and flashing, and only covers this model.
Linux firmware for the Echo Dot (2nd gen): no Fire OS, no Alexa. A local Home Assistant voice satellite with all seven mics, Bluetooth speaker mode, signed A/B updates and SSH.
★38
An Express middleware that caches responses by Idempotency-Key and replays duplicates, with memory or Redis storage, hitting the real pain of payment and order retries. Small in scope and accurately described, installable today; similar implementations exist, but this one works out of the box.
Express middleware that makes handlers safe to retry — replays the cached response for a duplicate Idempotency-Key instead of re-running it, with pluggable in-memory or Redis storage.
★17
A simulation environment, zero-shot scripts and trained checkpoints for dexterous-hand pen spinning and Rubik's cube solving, aimed at embodied AI researchers, and such reproducible in-hand manipulation resources are scarce. The limits are binding to a specific simulation stack and compute, making near-term deployment hard.
Dexterous pen spinning and Rubik's cube: sim environments, zero-shot scripts, trained checkpoints
★22
That's today. The other tiers and the boards follow.
Trains a 700K-parameter small model for the many local decisions in computer-use (checking boxes, skipping, filling values), pulling System1-specific judgment out of the general LLM loop at 7ms versus 260ms; backed by the 24K-star cua, open-source drivers and benchmarks, and HN's 56 points show agent vendors are genuinely watching; the risk is that jev-style services become the standard first.
Hello HN! We're Dillon and Francesco from Cua. We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their …
▲56
It packages HTML apps, assets and SQLite data into a single distributable file, solving the long-standing pain of static pages having nowhere to store data, with a permission model and local AI. 267 upvotes is the strongest momentum this period, and the 1.0 open format could become a standard; commercialization could be an app marketplace and enterprise distribution, with the risk of WASM sandbox alternatives.
Hey HN, I always had the problem that building HTML pages is really simple now, but trying to save data required hosting it somewhere, and sharing it afterwards was not easy. Over the last few months,…
▲267
Compresses tool calling and structured JSON output into 8–29MB 2-bit models that hit 4k tok/s on Raspberry Pi and browser WASM, cutting edge-agent deployment costs. The third iteration earned 156 votes, showing the team keeps delivering on feedback. It could license an edge inference SDK. Risk: narrow capability, squeezed by edge runtimes like llama.cpp and MLX.
Hey HN, Henry from Cactus here. We submitted Needle 2 here a few weeks ago, and the feedback in the discussion thread was incredibly valuable, thanks! Thanks to all that feedback, we’ve been able to m…
▲156
Post-training Qwen 27B with on-policy distillation that penalizes tokens tied to overthinking halves think length with almost no accuracy loss, directly cutting inference cost. It drew 80k downloads in 3 days plus community quantization and independent evals, strong momentum; monetization is API and licensing (only charging companies above $10M), but the risk is that base models' built-in effort control will absorb this value.
Hi everybody, we post-trained Qwen 3.8 27B to be more efficient by figuring out which tokens were linked to overthinking and penalizing them without "attacking" the reasoning length directly then fixe…
▲26
It rebuilds Qwen3-TTS/ASR into a speech-specific inference engine, proving open-source speech can be cheaper than closed source with sub-50ms latency and top WER on the Coval leaderboard, selling endpoints directly. 59 HN upvotes hit the cost pain point. Commercialization is the clearest here, a usage-based inference service; the risk is Alibaba's official optimizations erasing the price gap and vLLM adding multimodal inference.
Hey HN, Toby from Nari Labs here. We've been working on making OSS speech models super-fast. Last year, we built Dia, the first OSS text-to-speech model capable of doing natural dialogue. Since then, …
▲59
A Raspberry Pi e-ink photo frame identifies birds offline, then uses local AI to generate 19th-century hand-drawn style illustrations, grounding audio recognition and image generation in concrete hardware. 1237 upvotes and 1415 stars make it the strongest momentum this period. It could go consumer birdwatching hardware plus subscription against Bird Buddy; the risk is that recognition accuracy and cost determine repeat purchases.
See also: Avian Visitors - https://news.ycombinator.com/item?id=48343424 - May 2026 (20 comments)
▲1237
Turns long-running AI agents into an email-like self-hosted inbox, using MCP and SKILL.md to manage approvals, memory, and resumable interruptions, solving observability for unattended background tasks, validated by 2,000 people at Amazon. Monetization could be an enterprise edition and skill marketplace, but the risk is head-on collision with big-vendor desktop agents.
Hi HN - long-time lurker (since 2012!), first time poster. Pizza Bot is a self-hosted desktop app for Mac, Windows, and Linux that runs AI agents in the background and exposes them through an email-li…
▲17
A testing and evaluation platform for MCP servers, addressing the inability to observe MCP services in agent clients or run regression tests; the founder comes from Asana and brings an open-source user base as momentum. The commercialization path is clear: per-seat CI integration pricing, with the risk of being squeezed by Anthropic's official Inspector and various free evaluation tools.
Prathmesh, CEO of MCPJam here. Users now start in ChatGPT, Claude, Cursor, and other AI clients. They reach your product through your MCP server. That means your users often aren’t in your product any…
▲9
OpenRouter for agent tools routes and calls tools from different vendors through one interface, fixing the fragmentation of rewriting code for every tool added. 67 stars plus Show HN hit a real standardization pain point; the path is take rates and marketplace distribution, with potential to become a standard gateway, but the risk is MCP official and ecosystems like LangChain building in the same capability.
▲10
Makes a 500TB web index queryable as read-only SQL/Datalog, allocating compute via congestion bidding so agents escape uncontrollable black-box search APIs. 42 votes show the search-controllability pain is real. It could charge per query as agent-facing search infrastructure. The biggest risk is search-engine-scale cost, Exa and Tavily already occupying the space, and data coverage being hard to match.
Meet Scry, a 500 TB NVMe internet index in ClickHouse that you can run ~arbitrary readonly SQL and some of Datalog over, and I handle the problem of resource-contention with congestion-based micro-auc…
▲42