-
MIT-licensed 26M-parameter function-calling model distilled from Gemini 3.1 — runs tool calls on phones and watches at ~6000 tok/s prefill, with open weights and the dataset-generation pipeline published, not just the inference code. Maintained by Cactus Compute (the on-device-inference company), so the OSS release sits upstream of their commercial stack rather than being a teaser for it
-
GPL-3.0 browser extension that doesn't hide ads — it replaces them with OBEY / CONSUME / NO INDEPENDENT THOUGHT slogans from the 1988 film. A one-developer art project rather than a filter-list business, so there's no acquisition-and-enshittification path the way there is for commercial blocklists
-
GPL-licensed coding agent in pure Rust that holds an empty session in ~8 MB of RAM (~12 MB working), against the multiple gigabytes Claude Code and OpenCode grow into on large repos — Unix-pipe-shaped and single-author, with the bus-factor caveat that comes with a solo maintainer
-
AGPL-3.0 self-hosted "internet OS" (desktop, files, browser-based apps) out of beta after two years, 370 contributors and 40k stars — the contributor count, not the star count, is the signal it can outlive its founders. The hosted tier funds the open-source build
-
Closed-source merchant-of-record and billing platform (payments, global tax, subscriptions, usage-based) from the team behind ParityDeals — takes 3.5% + $0.40 per transaction (2.9% intro on the first $5K). The transaction cut is the funding model, the same shape as Paddle, Lemon Squeezy and Polar, in a consolidating MoR category where the open question is differentiation, not survival
-
Apache-2.0 Android screen mirroring and control over USB or TCP with nothing installed on the phone and no account or cloud relay — a major version bump on a decade-old Genymobile-maintained tool that has never had a paid tier to lose
-
AGPL-3.0 sync tool that pushes shared quality profiles and custom formats into your *arr stack, now able to connect multiple community config databases at once — its continuity question is whether those upstream databases stay maintained, not the app itself
-
MIT scripts that turn an $80 RK3562 Android tablet into a Debian Linux workstation — a single-author, hardware-specific hack, so treat it as a recipe to fork rather than a product with a roadmap
-
MIT CLI that auto-detects your GPU/CPU and ranks local LLMs by merged real benchmarks (LiveBench, Arena ELO, Artificial Analysis) with confidence-weighting that penalises stale numbers — instead of just matching model size to VRAM. Ranking quality decays if the upstream benchmark feeds aren't refreshed
-
Closed-source managed scraping API that turns any URL into JSON or Markdown for agents, handling proxies, rendering and retries — free 1,000 calls/mo, then $49–$249/mo, billed only on successful requests. Hosted-only and metered, so cost scales with crawl volume and there's no self-host escape hatch — the trade-off for not running your own proxy pool
-
MIT self-hosted podcast ad-stripper — Whisper transcribes, your own LLM finds the ad reads, FFmpeg cuts them, your app subscribes to the cleaned feed. New release adds opt-in crowdsourced sponsor patterns (off by default, nothing syncs unless you enable it). Bring-your-own-LLM, so inference cost is yours, not the project's
-
MIT code-search library for agents that answers natural-language queries in ~1.5 ms on CPU using roughly 98% fewer tokens than grep-plus-read (≈2k vs ≈100k tokens for comparable recall) — a library, not a service, so there's no API to meter or rate-limit later
-
Apache-2.0, S3-agnostic self-hosted WeTransfer/Dropbox alternative, now with no-account quick-share and reverse-share links (expiry, view caps, password, per-file size limits) — a two-person project still early at v0.5, so the sustainability question is bus-factor, not monetization