The Open Source Drop

Two tools most of us stopped reading release notes for changed this year, and neither made much noise about it. Transformers, the library behind almost every AI tutorial you have read, dropped TensorFlow and JAX in version 5. It is PyTorch only now, so a lot of older example code no longer runs as written. VS Code folded Copilot Chat into the main repo as open source and built AI into the editor itself, with a single setting that turns all of it off. When roughly three out of four developers use the same editor, that setting is worth knowing about. Defaults get to do that. Once a tool is the thing nobody has to justify choosing, it can change underneath you and you find out from a broken tutorial. The other two picks are trying to earn that position, and both are doing it the same way: by removing the setup. DSH Desktop, this week's featured tool, wraps DeepSeek's agent runtime in a plain installer for Windows and Mac, so trying it no longer starts with a terminal. It is a community project, not DeepSeek's, and it is free. Understand-Anything does the same for your own code. It maps a repo into a knowledge graph you can click through and question, then saves the graph as JSON so the next person skips the expensive part. One honest warning on the featured pick. DSH Desktop pins a release candidate of a runtime that calls itself a developer preview and promises breaking changes in capital letters. Install it to learn the runtime. Keep it off anything with a deadline.

September 22, 2026

Free is doing a lot of work in this issue. Four open source tools, nothing to pay to download, and that is where the resemblance ends. SGLang, the featured pick, is the cleanest case: Apache 2.0, no strings, serving open-weight models from one GPU up to a cluster. The price is operational. Cold starts and cache tuning are still yours. commerce-agents is free the way a blueprint is free. Anthropic published runnable reference agents with the safety scaffolding already wired in. The code costs nothing. Claude does not. OpenWhispr flips it again: hit a hotkey, talk, and it types wherever you are, transcribing locally with no API key and no audio leaving the machine. Free forever for one person, a subscription once a team wants sync. Recordly is the one to read the fine print on. AGPLv3, plus two terms most people will not expect: no use of the Recordly name, and credit in your user-facing UI if you build on the code. Record your own demos and none of it touches you. Ship something built on it and all of it does.

September 15, 2026

Asking one model for one answer is starting to look like the least interesting thing you can do with AI. Every tool in this issue works the other way: run several attempts, let them disagree, keep the one that survives. OpenMAIC is the clearest version and it is this week's featured pick. It builds a classroom where AI teachers lecture and AI peers discuss and push back, so what you walk away with came out of the exchange rather than a single flat answer. It came out of Tsinghua, it is grounded in a paper in the Journal of Computer Science and Technology, and it is free to run. Orca does the same thing to your codebase. Fan one prompt across five agents, Claude Code and Codex and OpenCode and whatever else you already pay for, each working in its own isolated git worktree, then read the diffs side by side and merge the one that actually got it right. Marin applies the idea to training itself: an open recipe for building foundation models where every checkpoint, data mixture, and decision gets published, failed experiments included. That last part is rarer than it should be. Most labs publish the run that worked and quietly bury the nine that did not. OmniVoice Studio is the outlier and earns its spot differently. It clones a voice from a three second clip, dubs video end to end, and builds audiobooks out of EPUBs, all on hardware you own instead of somebody's API. One thing to check before you build on it: OmniVoice is AGPL-3.0, so if you modify it and offer it to other people as a service, you owe them your changes. OpenMAIC and Orca are MIT, Marin is Apache 2.0. Only one of the four bites you on licensing, and it is the one you would most want to wrap in a product.

September 8, 2026

One developer ran 740 job listings through career-ops and came out the other side with a Head of Applied AI role. The part worth stealing is not that an AI wrote the resumes. It is that somebody looked at a job search, which is normally a shapeless pile of tabs and half-remembered conversations, and built a pipeline for it: every listing broken into a structured A through H report with a single 1 to 5 score on top, an interview story bank in STAR format, a terminal dashboard to see the whole funnel. Structure beat volume. That is the thread running through the whole issue. OpenViking organizes an agent's memory the way a filesystem organizes files, directories it browses with ls, tree, and find, instead of the usual bag of vectors you throw a query at and hope. Every retrieval leaves a trajectory you can watch, so when your agent recalls the wrong thing you can see exactly which path produced it. DeepSeek Harness, this week's featured pick, takes that instinct down to the architecture: everything is a plugin, built on top of Cordis, so pieces get swapped rather than forked. Plane does it for people instead of agents, cycles and roadmaps and a snappy UI, minus the Jira bill. Two things to check before you commit to any of it. Plane is AGPL-3.0 and so is OpenViking's core, though OpenViking's CLI and examples are Apache 2.0, so know which piece you are touching. AGPL means that if you modify it and offer it to other people as a service, you owe them your changes. Fine for an internal tool, a real problem if you are building a product on top. And DeepSeek writes it in capital letters in their own README: Harness is a developer preview and there will be compatibility-breaking changes. The plugin API you build against this month may not be there next month. Worth an evening, not worth a roadmap.

September 2, 2026

I keep a running list of things I was told needed a datacenter. Fine-tuning a model sat on it for a long time. Unsloth, this week's featured pick, cuts the VRAM requirement by about 70% and roughly doubles training speed, which is the difference between renting a datacenter GPU by the hour and using the card already sitting in your desktop. That line moved, and it moved a lot. It moved on the inference side too. oMLX puts a local OpenAI-compatible server behind a Mac menu bar icon, so Claude Code or Codex points at localhost instead of an API that bills per token. And antirez, the guy who wrote Redis, published a from-scratch C engine that generates video with synchronized audio on a Mac. No cloud, no API key, and the code is short enough to actually read. That last part matters more than it sounds. Most of this stack is impenetrable Python, and a readable C implementation is how people figure out what is really happening underneath. ToolJet is the outlier this week, a self-hosted Retool alternative for building admin panels, and it fits the pattern anyway. Everything here runs on a machine you control. That was not a reasonable expectation eighteen months ago.

August 26, 2026

Every AI project hits the same wall in week one: the data you need is trapped in Word docs, PDFs, and somebody's Excel export. anydoc, this issue's featured tool, is Firecrawl's converter for exactly that. Word, PowerPoint, Excel, PDF, EPUB, all of it comes out as clean Markdown in milliseconds, and it's the fastest-moving project we're tracking right now. One catch worth knowing up front: it reads the text layer of a PDF, not scans. If your PDFs are photographs of paper, you still need OCR first. The rest of the issue is agents growing up. Comp AI built a CRM where the agent does the work on its own schedule and the CRM is just where it takes notes. qm is Y Combinator's own system for running a whole team's agents, one isolated workspace per person, driven from Slack or the web. The pattern is hard to miss: agents are moving out of demos and into org charts. And because somebody has to keep the lights on, Authentik rounds it out: self-hosted single sign-on and multi-factor auth, the stuff you'd otherwise buy from Okta or Auth0. Less exciting than the agent stuff, and probably the one your company actually needs first.

August 18, 2026

Three of the four tools this week replace something you're probably paying for right now. That wasn't planned, it just came out that way, and it says something about where open source has gotten to. Kaneo is the featured pick: a kanban board and task tracker that skips the enterprise bloat. Free, self-hosted, and deliberately not trying to grow into Jira. Quill does meeting recording and transcription entirely on your Mac, nothing uploaded, which is the exact opposite of how Otter.ai and Granola work. Remotion turns video into React components, so a data-driven animation becomes JSX instead of an After Effects template. Free under $100K in revenue, then you pay, and that's a fair line to draw. AgentENV is the outlier and worth a look anyway. It runs thousands of isolated agent environments on Firecracker microVMs for reinforcement-learning training, and it can pause and resume them in under 50 milliseconds. Most of you will never need it. If you're training agents at scale, it's the infrastructure that makes it practical.

August 11, 2026

The most interesting tool this week is an agent skill called img2threejs, and the smart part isn't what you'd guess. It takes a reference photo and rebuilds the object as procedural 3D code you can animate. But the model barely does the heavy lifting. Deterministic Python, standard library only, handles the validation and the code generation. The AI's tokens go only to the visual judgment: look at the image, assess it, check the render. That split is why it stays cheap to run, and it's a pattern more agent skills should copy. That's the thread running through this whole issue. Agents are moving out of the chat window and into real production work. Palmier Pro is a native Mac video editor that hooks into Claude, Codex, and Cursor over MCP, so an assistant can edit the timeline alongside you. RAGFlow grew from a document-parsing engine into a full agent platform sitting on your own files. LangGraph is the framework people reach for when an agent workflow stops being a straight line and starts needing loops, branches, and approval steps. None of these are magic. Palmier's editor is free and open source, but the in-timeline AI generation is the paid, closed part, so know what you're getting. RAGFlow self-hosted is a real ops project, not a docker run. And img2threejs is honest about its own limits: it's strongest on hard-surface objects and tells you when it can't hit the fidelity you want from one photo. That kind of honesty is rare, and it's worth rewarding.

August 4, 2026

Ask your AI assistant what time it is and it will make something up. Models don't have a clock. The featured pick this week is the official fix: a tiny MCP server that gives your assistant the current time in any timezone, running locally with no API keys. It's the least flashy thing on this list, and it might be the one you end up using every day, because scheduling, logs, and cron all fall apart when the model can't tell Tuesday from Thursday. The other three swing the opposite direction. T3MP3ST points a swarm of AI agents at a target, tries to break in, and writes up how it got there. It doesn't even want its own API keys, it drives the coding agents you already run. SigNoz puts traces, metrics, and logs in one OpenTelemetry-native dashboard, the observability stack you'd otherwise pay Datadog for. And Home Assistant is the biggest name in home automation for a reason: one dashboard for every light, lock, and thermostat you own, running on your own hardware. That's the spread this week: a one-job utility, an autonomous red team, a full observability stack, and your entire house. Open source doesn't really have a lane anymore.

July 30, 2026

Harvard put a full textbook on ML systems online for free, and it's the featured pick this week. Not something you install. Something you read, at mlsysbook.ai. It covers the layer most of us skip: hardware, quantization, deployment on tiny devices, what actually happens underneath the model. The industry is full of people (me included) who learned AI by calling APIs. This is the book for closing that gap, and it costs nothing. The other three are things you run. QwenPaw is a personal AI assistant that lives on your own hardware and talks to you through Discord, Telegram, even iMessage, no API key needed out of the box. Dify keeps showing up in production AI stacks: a workflow editor and RAG pipeline you can self-host on Docker. And Logto wants to be the auth system you own instead of rent, with SSO, MFA, and user management and no feature gates on the self-hosted version. A textbook, an assistant, a platform, and an auth layer. No grand theme this week, just four things worth your time.

July 14, 2026

The models get the headlines. The tooling around them is where the real work is happening right now, and this week made that obvious. Baidu dropped an OCR model under an MIT license that reads an entire PDF in one pass, no page-by-page babysitting, free even for commercial use. That's the featured pick. The fact that you can run document extraction like that for nothing still gets me. The catch is the GPU, but the ceiling on what's free keeps climbing. The other three live in that same layer. vLLM is the engine a lot of people are quietly using to serve open models. oh-my-openagent splits coding work across a crew of sub-agents instead of betting everything on one. And Cloudflare shipped a security skill that turns your coding agent into an auditor that only reports what it can actually exploit, which is the discipline this space has been missing. Four tools, one theme: the plumbing is getting really good.

July 7, 2026

The project everyone's piling onto right now isn't a database or a framework. It's ponytail, a skill that does one stubborn thing: it tells your AI coding agent to write less code. Reach for the standard library. Use what's already there. Stop reinventing what the language already does. Anyone who's watched an agent bloat a file with code that didn't need to exist knows exactly why people want this. The rest of the list rhymes with that instinct: keep control of your own stuff. Penpot is the closest thing open source has to Figma, and it puts your design files on your server instead of someone else's roadmap. Turso is SQLite pushed out to the edge, close to your users, free to self-host. World Monitor is an OSINT-style dashboard that plots global events on a 3D globe and can run its AI analysis locally through Ollama, so nothing has to leave your machine. None of these are perfect. Ponytail is a nudge, not a guarantee. Penpot is not Figma. Turso is SQLite, not Postgres. We say so in every writeup, because the catch matters as much as the pitch. Here's what caught our eye this week.

June 30, 2026

Two of this week's tools install with a single npx command and then live inside Claude Code. One turns a text prompt into a Lottie animation. The other lays out a resume or a slide deck that doesn't look like a Canva template. You're not running an app, you're handing your coding agent a new trick and letting it do design work it had no business doing a year ago. That document designer is named Kami, which also happens to be my three-year-old daughter's name, so I'll be straight with you: it was making this list either way. It earned the spot on its own, but I wasn't going to pretend to be neutral about it. Chatwoot is the featured pick and the most production-ready of the bunch: a self-hosted support inbox that covers the core of what Intercom and Zendesk bill you per agent for. The honest catch is operational. You're standing up Rails, Postgres, Redis, and Sidekiq, so it fits if you've got the ops muscle and want off the per-seat treadmill, not if you want something running by lunch. openhuman is the long shot. A local-first personal assistant that keeps your data as plain markdown on your own machine and connects to a hundred-plus services. It's early beta and says so plainly. Worth a look if you like poking at ambitious privacy-first software, worth skipping if you need it to just work.

June 23, 2026

Four tools caught my eye this week, and they all point the same direction: stop renting your AI assistant from a cloud vendor and run it on your own hardware. The privacy pitch is finally backed by tooling that doesn't take a PhD to set up. odysseus is the one everyone's talking about. It's a self-hosted AI workspace that bundles chat, autonomous agents, email triage, and research into one app, and it can serve open models directly. The clever part is the Cookbook: it looks at your actual hardware and recommends models you can really run, then wires them up for you. It launched to a massive audience overnight, so expect rough edges. And read the security note first. The project itself tells you to treat it like an admin console and never put it on a public IP. The other three fill in the stack. supermemory gives your agent a long-term memory that survives across conversations. Open WebUI is a polished ChatGPT-style interface for whatever models you're running locally. DeepSeek-GUI puts a real review step in front of an AI coding agent, so you see the diffs before anything changes. The common thread: you own your data, and you pay for it in hardware and setup time instead of a monthly bill. Whether that trade is worth it comes down to what you've got sitting under your desk.

June 16, 2026

This week kept circling back to one question: what does it actually cost to run an AI agent, and who's doing something about it? Three of the four tools below are different answers. headroom is the one I keep thinking about. It sits in front of your LLM calls and strips the junk out of everything your agent reads (logs, tool output, RAG chunks, file dumps) before it ever hits the prompt. The reported numbers hold up: 87% fewer tokens on a log-search test, accuracy unchanged. One command to install, runs entirely on your machine, nothing leaves your laptop. If you're running coding agents hard, it pays for itself almost immediately. The rest follow the same instinct: own your stack, control your costs. DeepSeek-Reasonix is a terminal coding agent built around prefix caching, so a session that would run sixty dollars comes in closer to twelve. Hermes WebUI gives you a self-hosted agent with memory that isn't chained to ChatGPT or Claude.ai. And Twenty is the outlier, a Salesforce replacement that actually looks like it was built this decade. Four tools, one theme: keep your data and your bill under your own roof.

June 9, 2026

Something shifted this week. The tools that stood out weren't built for developers. They were built for the agents developers run. Microsoft shipped a governance layer that sits between your AI agents and the actions they take, evaluating every tool call against policy before it executes. Vercel Labs went further and released a programming language whose intended reader isn't a person at all, it's the compiler talking back to a model in clean JSON. Different problems, same premise: agents are now first-class users of our tooling, and the tooling is starting to assume it. Not everything went that way. n8n is still doing the unglamorous work of wiring 400 services together so you don't have to, and html-anything just borrows the agent CLI you already pay for to turn notes into polished HTML. But the direction of travel is hard to miss. We spent years making developers faster. Now the tools are aimed one level down, at the agents doing the work.

June 2, 2026

Bun is the one I keep coming back to. It's a runtime, a bundler, a test runner, and a package manager in a single binary, and it's faster than the Node tools it replaces at nearly every one of those jobs. The whole thing is free, no paid tier hiding the good parts. For a new project, there's almost no reason not to try it. The rest of this week leans the same direction: tools that let you own the thing instead of renting it. Plausible is a clean, privacy-first analytics dashboard you can self-host, no cookie banner, no handing your visitor data to Google. Cal.com does what Calendly does, scheduling links and calendar sync, except you can run the whole thing on your own server. And Lark's official CLI wraps 200+ commands across their suite, so if you live in Lark you can automate most of it from a terminal. None of these are new ideas. What changed is how good the open source versions got. A few years ago, self-hosting your analytics or your scheduling meant accepting a worse product to save money. That trade-off is mostly gone. Pick the tool because it's better, not because it's cheaper.

May 26, 2026

Four projects caught attention this week, and they share a posture: each does one thing well and refuses to become a platform. Browser Harness gives LLMs raw Chrome control through a single WebSocket in 600 lines of Python. whatcable surfaces the USB-C cable specs macOS hides from you. Open Design generates websites, mobile apps, and decks by delegating to whichever AI CLI you already have. Floci is a drop-in LocalStack replacement that boots in 24 milliseconds. None of these are trying to be your operating system. Worth calling out: LocalStack ended its free community tier in March, and the OSS replacement story has been settling out over the past two months. Floci looks like the one that stuck.

May 19, 2026

Four tools this week and they all have one thing in common: they're trying to put AI inside the apps you already live in. Warp is rebuilding the terminal around it. Zed is doing the same thing for the editor. Open CoDesign turns prompts into UI prototypes on your desktop. openclaw stitches every messaging app you use to one local AI assistant. The pattern that keeps showing up is bring-your-own-key. Three of these four are MIT-licensed and ship with no model provider lock-in. You point them at OpenAI, Anthropic, Ollama, whatever you want. "Open source AI tooling" increasingly means "we built the interface, you pay your own model bill," and the economics line up. Nobody is trying to resell you tokens at a markup. The exception is Warp. It's source-available, not open source. The license restricts what you can do with the code and the AI features require an account. Worth flagging because the polish is real but the freedom is partial. If that bothers you, Ghostty and Alacritty are the open answers, and they're in the database too.

May 12, 2026

AI agents stopped being a single product. We have agents in our terminals, our chat apps, our IDEs, our browsers, and now we have launchers to switch between them. Three of the four tools in this issue are agents. The fourth uses one as a building block. Hermes is the most interesting one to me. Nous Research built it as a self-improving agent that runs in your terminal but also plugs into Telegram, Discord, Slack, WhatsApp, Signal, and email. Point it at any LLM provider you want, and it remembers what you taught it across sessions. opencode is the same energy for code: terminal-first, model-agnostic, MIT licensed. cc-switch exists because if you've installed Claude Code, Codex, OpenCode, openclaw, and Gemini CLI on the same machine, you need something to keep them straight. The sneaky one is llm_wiki. Drop a folder of PDFs and Word files into it and it builds a structured wiki with cross-references and a knowledge graph, using whatever LLM you point at it. Not an agent in the chat sense, but the same pattern applied to documents instead of conversations. Persistent state, built up incrementally, owned by you. The open source version of the AI tools we used to subscribe to is starting to feel like a coherent stack instead of a collection.

May 5, 2026

Something happened this week that I haven't seen before. A tool that literally just tells Claude Code to stop being polite blew up faster than anything in our database. Caveman doesn't add features. It removes words. And developers are installing it by the thousands because apparently we all wanted our AI to cut the "I'd be happy to help" and just do the work. Separately, Graphify caught my attention because it solves a problem I hit constantly: understanding a codebase you didn't write. It reads your code, docs, even screenshots, and builds a navigable knowledge graph. Not a summary. Not a chatbot. An actual graph you can explore. That's a different approach and I think it's the right one for complex projects. Also on radar this week: an agent framework shipping 43 built-in tools out of the box, and a Rust-based S3 alternative that's picking up serious momentum in the self-hosted storage space.

April 21, 2026

Claude Code has an ecosystem now. Not just plugins or extensions, but a full configuration playbook: skills, memory structures, security practices, and CLAUDE.md templates that change how the agent behaves. everything-claude-code is the fastest-growing project in our database this week, and it's not hard to see why. Developers are realizing the default setup leaves performance on the table. Also on radar: a Rust-based terminal agent that replicates Claude Code's workflow for free, a proxy that turns AI coding CLIs into standard API endpoints, and Immich, the self-hosted photo platform that keeps pulling people off Google Photos.

April 14, 2026

The Number: 66% of the tools we track are completely free — no paid tier, no catch.

April 7, 2026

The Number: 66% of the tools we track are completely free — no paid tier, no catch.

April 1, 2026

Welcome to The Open Source Drop, a free, no-BS look at open source tools worth knowing about. Every issue: a few tools we've actually researched, with honest analysis of what they do well and where they fall short. No sponsored picks. No hype.

March 24, 2026