Open Source Radar — September 1: agents that finish the job
Today's daily board has a clear pattern: agents are no longer just writing code in a terminal, they're producing finished artifacts — a paper, a cut video, a cloned voice, a model of your own. Five fresh projects, none of them repeats from this week's earlier radars.
openclaude (TypeScript, ~30,900 stars) — A terminal coding agent that refuses to care which model you point it at. One CLI covers OpenAI-compatible APIs, Gemini, GitHub Models, Codex OAuth, Ollama and others, with a guided setup that saves provider profiles so you can switch backends without rewriting config. It carries the full agent toolkit — shell, file tools, grep, agents, MCP, slash commands — plus a structural repo map ranked by importance that it injects into context so the agent knows your codebase before it starts editing. Background sessions let you fire a long task, detach, and check on it later, and there's a bundled VS Code extension for people who want the chat pane. The interesting part is the license-to-model split: keep the workflow, change the brain. You'd reach for it if you're tired of maintaining a separate toolchain per provider.
academic-research-skills (Python, ~44,300 stars) — Four Claude Code skills that run the whole academic pipeline: research, write, review, revise. Under the hood they're multi-agent teams — 13 agents on deep research, 12 on paper writing, 7 on review — with modes for literature reviews, systematic reviews, rebuttal audits, and citation checks, and output that lands as Markdown, DOCX, and LaTeX (APA, IEEE, Chicago) rendered to PDF. The maintainer publishes the real artifacts the pipeline produced, including a post-publication audit that caught 21 issues three rounds of peer review missed, which is a more convincing demo than a screenshot. Fair warning on the license: attribution required, commercial use not allowed, so this is for researchers and students rather than a company build.
minimind (Python, ~56,200 stars) — The from-scratch LLM teaching repo that keeps climbing, now past fifty-six thousand stars. Its claim is still the whole pitch: train a 64M-parameter language model in about two hours on a single consumer GPU (a 3090), which the author prices at roughly three yuan in rented GPU time. It walks the full modern curriculum — pretraining, supervised fine-tuning, LoRA, tool calling, GRPO-style reinforcement learning, and agentic RL — in code small enough to read in an afternoon, and the trained weights drop into llama.cpp, vLLM, Ollama, or SGLang for inference. It's the fastest way to understand what a model actually is by building one, and the tokenizer is tuned for mixed Chinese-English text, which is why it has such a following in China's LLM-learning community.
video-use (Python, ~22,500 stars) — Browser-use's answer to "agents can write code, so why not cut video?" You drop raw footage in a folder, tell your coding agent what you want, and it returns a rendered file: filler words and dead air removed, every segment color-graded, 30-millisecond audio fades so cuts don't pop, subtitles burned in, and animation overlays generated in parallel sub-agents. The detail that makes it more than a script is that the agent evaluates its own render at every cut boundary before showing you anything, and keeps a session memory file so next week's edit picks up where you left off. It's MIT-licensed, works under any agent with shell access, and needs ffmpeg plus an ElevenLabs key for the voice work.
VoiceStudio (Python, ~12,900 stars) — A fully local alternative to ElevenLabs: voice cloning, voice design, video dubbing, dictation, transcription, and audiobook generation, running on your own machine with no account, no API key, and no usage meter on the core workflow. It fronts 16 text-to-speech engines and 11 speech-to-text engines behind one desktop app, ships a local REST and MCP server so an agent can drive it, and covers a claimed 646-language catalogue — coverage and quality depend on the engine you pick, which is the honest caveat. It runs on CUDA, Apple Silicon, ROCm, or plain CPU, and it's AGPL-licensed, so check the terms before embedding it in a product. Worth a look if your audio pipeline currently sends every file to someone else's cloud.
Worth watching this week: the agent stack is moving from "writes a diff" to "delivers a file" — once agents can hand you a paper, a video, or a model, the bottleneck shifts to whether you trust the output enough to ship it.
Which of these would you hand a real deadline to — the paper, the video, the voice, or the model? Tell us in the comments.
Sources: openclaude (GitHub) · academic-research-skills (GitHub) · academic-research-skills architecture docs · minimind (GitHub) · video-use (GitHub) · VoiceStudio (GitHub)
