Open Source Watchlist
Twenty repos worth watching. Organized by category. Not ranked — we don't do rankings. These are things we've actually run, used, or studied closely enough to have an opinion on. Stars are approximate and shift constantly; treat them as a rough signal of community size, not quality.
01 AI / ML Tooling
The tooling layer around LLMs — inference, evaluation, orchestration. Moving fast. Most of these didn't exist two years ago.
Unified interface to call 100+ LLM APIs using the OpenAI format. Supports routing, fallbacks, spend tracking, and caching. Works as a proxy or SDK.
The router logic alone is worth it. We use this in production to failover between providers without rewriting call sites.
Open source LLM app development platform. Visual workflow builder, RAG pipeline, model management, and observability in one self-hostable package.
More polished than most no-code AI tools, but still has rough edges if you go off the happy path. The RAG pipeline is genuinely good.
LLM evaluation framework with a Pytest-like interface. Covers hallucination, answer relevancy, contextual precision, bias, and toxicity metrics.
Eval tooling is still immature but this is the least painful option we've found. The pytest integration makes it easy to add evals to CI.
CLI tool and Python library by Simon Willison for running prompts against local and remote LLMs. Plugin system extends it to dozens of models.
Unreasonably useful for quick experiments. The logging to SQLite is a detail that saves you constantly when you need to reconstruct what you ran.
02 CLI Tools
Command-line tools that earn their place in a workflow. We're skeptical of rewrites-in-Rust that don't actually improve on the original, so everything here has cleared that bar.
Smarter cd command. Learns your most-visited directories and lets you jump with partial matches. Written in Rust. Integrates with bash, zsh, fish, nushell.
After two weeks you forget it's not just how cd works. The fzf integration turns it into a fuzzy directory picker which is where it gets genuinely fast.
cat clone with syntax highlighting, Git integration, and automatic paging. Supports 200+ languages. Drop-in replacement for most cat use cases.
This one actually replaced cat for us. The Git diff sidebar is the feature nobody talks about but uses constantly once they've seen it.
Markdown renderer for the terminal. Supports stashing, paging, and rendering from stdin, local files, or URLs. From the Charm ecosystem.
The stash feature is underrated — treating local markdown as a document library is genuinely useful for reading READMEs without opening a browser.
Modern HTTP client for the command line. Human-friendly syntax, JSON support, sessions, plugins. curl replacement for API testing and one-off requests.
We reach for this when curl flags start looking like line noise. The session support is what separates it from just being a syntax wrapper.
03 Dev Workflow
Tools that change how you work, not just what you use. Workflow tools earn their place slowly — we've lived with all of these for at least three months.
Terminal UI for Git. Interactive staging, branch management, rebasing, and conflict resolution without leaving the terminal. Written in Go.
The interactive rebase view alone is worth the install. We've watched multiple Git CLI veterans switch to this and not go back within two weeks.
Command runner for project-specific tasks. Like make but without the build system baggage. Justfiles are readable, tab-complete-friendly, and cross-platform.
We've standardized on this across all projects as the entry point. New contributor? just --list. The self-documentation alone justifies the choice.
Syntax-highlighting pager for Git diffs. Configurable themes, side-by-side view, line numbers, merge conflict resolution display. Drop-in via core.pager.
Sets a new floor for what diff output should look like. After configuring this, reviewing diffs in the browser feels like a regression.
Structural diff tool that understands syntax. Compares code structurally using tree-sitter parsers rather than line-by-line. Supports 30+ languages.
Genuinely better for reviewing refactors where lines moved. The tradeoff is slower on large files — we use it for code review, delta for git log.
04 Self-Hosted
Infrastructure and services you can run yourself. We bias toward things that are stable, have real documentation, and don't require a PhD to operate.
Self-hosted Heroku/Netlify/Vercel alternative. Deploys apps, databases, and services from Git. Docker and Nixpacks support. Runs on any VPS.
Finally stable enough for production. Version 4 fixed the database migration pain that plagued earlier releases. We run it on a Hetzner VPS for non-critical projects.
Web server with automatic HTTPS via Let's Encrypt. Caddyfile config is dramatically simpler than nginx. HTTP/3, reverse proxy, load balancer built in.
Automatic TLS has been reliable for us across dozens of domains. We haven't touched an nginx config in over a year. The Caddyfile format feels almost too easy.
Self-hosted photo and video backup solution. Mobile apps, machine learning-based face recognition, albums, maps view, and partner sharing. Docker deploy.
The ML search is actually good. "Photos from hiking trips" works. We include this as a showcase example when arguing that self-hosted doesn't mean worse UX.
Self-hosted media server and download manager for anime. Integrates with AniList, torrent clients, and local file management. Built with Go and React.
Listed here for the architecture, not the use case. The way it handles metadata sync and local state is worth studying even if you never run it.
05 Interesting Experiments
Things that aren't production-ready or aren't meant to be, but represent ideas worth tracking. We don't recommend shipping these. We do recommend reading them.
High-performance UI layout library written in C. Inspired by CSS flexbox but with zero dependencies and designed for game engines and custom renderers.
Not for web projects. Worth reading as an education in how layout engines actually work at the memory level. The README is unusually good.
Reimplementation of SQLite in Rust with async I/O support. Aims to be an embeddable database compatible with the SQLite file format and wire protocol.
Not production-ready. The async-native approach is the interesting part — if this stabilises it could change how embedded databases work in async Rust apps.
High-performance collaborative code editor built in Rust using GPUI. Real-time multiplayer editing, AI integration, and a custom rendering stack. macOS and Linux.
The GPUI architecture is the real experiment here. We're not recommending it over Neovim or Cursor yet, but the performance ceiling is genuinely different.
Runs unmodified CUDA applications on AMD and Intel GPUs via a drop-in library replacement. Experimental, incomplete support, but makes CUDA portability real.
The existence of this is more interesting than the implementation. If it matures, CUDA lock-in becomes a much weaker argument for Nvidia-only stacks.