The shortlist
9 projects
The wider reading list: projects that stood out across the week, in their own words. Positions reflect the shortlist, not GitHub rank.
Also worth watching
Fast risers from off the boards — worth a look anyway.
The wider reading list: projects that stood out across the week, in their own words. Positions reflect the shortlist, not GitHub rank.
Fast risers from off the boards — worth a look anyway.
The weekly Star
A project by debpalash
Desktop app for local voice AI workflows: voice cloning and design, video dubbing, dictation, and long-form audio production.
ElevenLabs sells a simple deal: send your audio to its servers, pay by the character, get a good voice back. VoiceStudio offers the other one. The models run on your machine, there is no account, and nothing meters the work. You pay in hardware, a 2.3 GB first download, and some attention to licences.
It is a desktop app by Palash Debnath, formerly OmniVoice-Studio, and this is its third week on our boards. We first recorded it on September 3 at 15,355 stars; by September 21 it had 33,816. This week it sat fourth on the daily board twice and held a monthly-board spot every day. It also shipped three releases in three days, 0.5.4 to 0.5.6, just after retiring its Tauri shell for Electron. That is a project moving fast, which is both the reason to look and the reason to pin a version.
Stars at each capture
Rank on each board
Most local voice projects do one job. VoiceStudio puts several behind one window: cloning from a short sample, designing a voice from a description, dubbing a video, a floating dictation widget, and long-form work such as audiobooks and batch queues. Underneath sits a model catalogue of more than a dozen speech engines and about ten transcription engines, switched from the status bar with Cmd/Ctrl+E.
Dubbing is where the design shows. The pipeline is transcribe, translate, speak, and every stage can be swapped. Two translation engines, Argos and NLLB-200, ship with every build and run offline; the online ones are optional. If you already have a translation, from a person or anywhere else, Paste Translation maps it onto the existing segments without transcribing again.
Then comes the step most dubbing demos skip. Before any GPU time is spent, each translated line gets a predicted speaking time against its slot. Lines that will need an audible speed-up are marked Tight fit; lines that cannot fit at all are marked Won't fit with the overrun in seconds. You shorten the sentence first, instead of hearing a voice sprint through it later.
For cloning, the project asks for 5 to 15 seconds of clean speech. Given a longer clip, the default engine picks the 15-second passage with the most speech in it.
The application is AGPL-3.0. The voice you hear by default is a separate question. The default engine runs k2-fsa's OmniVoice, whose model card puts the code under Apache 2.0 and the pretrained weights under CC-BY-NC, because of the data it was trained on. VoiceStudio's own licence notice says a commercial licence for the app does not replace those terms.
So an audiobook for your own shelf is fine on the defaults. A client video or a monetised channel needs an engine whose weights allow commercial use, and that means reading each engine's page before you switch to it. The app makes switching easy. It cannot make the choice for you.
The same care applies to whose voice it is. The project asks you to clone only with permission, and it can add an invisible AudioSeal watermark to generated audio. The setting is under Settings → Privacy; check it before anything leaves your machine.
The backend mounts an MCP server at http://localhost:3900/mcp/, with
generate_speech, clone_voice and transcribe among its tools. A coding
agent can narrate a changelog or transcribe a meeting recording without
the audio going to a cloud API. Set two variables before you connect one:
OMNIVOICE_MCP_BASE_PATH=/path/to/one/folder
OMNIVOICE_MCP_OUTPUT_MODE=filesThe base path is the boundary for every file the server reads or writes; without it, path arguments are refused. Files mode returns a path instead of a WAV encoded as base64, which would otherwise land in the agent's context a megabyte at a time.
Intel Macs cannot run the local backend. Linux needs x86_64 with glibc 2.39 or newer, and the Docker images are amd64 only. On a 16 GB machine, dubbing swaps the speech and transcription models in and out of memory, so expect it to be slower there than the demos suggest.
We have not put VoiceStudio through a real job yet, so this is an editor's pick, not a review. A fair first test is a two-minute recording of your own, dubbed into a language you can judge, with the offline engines only. Count the lines flagged before generation, then count the edits you still make after it. Builds are on the releases page; start with Voice cloning and fifteen clean seconds of your own voice.
AI-powered CLI code review tool that reads Git diffs or whole files, queries a configurable LLM agent, and returns line-precise comments.
Days on board: 6 daily · 6 weekly · 2 monthly
An agent skill that runs a structured six-phase security audit of a codebase with adversarial validation and coverage tracking.
Days on board: 4 daily · 0 weekly · 2 monthly
Open business management platform combining ERP, CRM, HRM, ATS, project management, and employee time tracking for companies and teams.
Days on board: 5 daily · 0 weekly · 0 monthly
Pure-C inference engine that runs very large mixture-of-experts models across VRAM, RAM, and storage tiers on consumer hardware.
Days on board: 5 daily · 1 weekly · 0 monthly
Local-first workspace that turns coding agents into research agents for literature review, experiments, and research artifacts.
Days on board: 4 daily · 0 weekly · 0 monthly
Desktop app for local voice AI workflows: voice cloning and design, video dubbing, dictation, and long-form audio production.
Days on board: 3 daily · 0 weekly · 7 monthly
Browser-based 3D globe app that visualizes live public data—aircraft, ships, satellites, earthquakes, and cameras—with voice control.
Days on board: 1 daily · 7 weekly · 0 monthly
LLM-powered knowledge framework for enterprise document understanding, RAG Q&A, agent reasoning, and auto-built wiki knowledge bases.
Days on board: 2 daily · 7 weekly · 2 monthly
A Claude Code plugin billed as an agent harness operating system, with an optional hosted GitHub App for private repositories.
Days on board: 4 daily · 7 weekly · 0 monthly · specialist pick
browser-use, an established name in web agents, ships its own Jev play — the fastest and cheapest web agent — the wave is reaching the incumbents, still entirely off trending.
The Jev idea you can use today: a Claude Code plugin that scores every tool call in one fast request, drops or truncates the stale, and keeps what survives verbatim — compaction without the summary loss.