Open Source Radar — September 29: self-hosting gets specific
Today's four all bet on running the thing yourself — a local voice stack, a household assistant, your own cloud box for coding agents, and a weight-level answer to model refusals.
VoiceStudio (Python, AGPL-3.0, ~46,000 stars) — An open-source, fully local ElevenLabs replacement: voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation across 646 languages, with the models running on your own machine instead of in someone else's cloud. It hit today's trending page and the velocity is real — the repository has gathered roughly 46,000 stars since April. Two details matter more than the feature list: it exposes a local API and an MCP server, so a coding agent can call it directly, and it ships as a desktop app rather than a research script. The catch is licensing. The code is AGPL-3.0, but each speech model carries its own terms — read them before anything commercial, and clone only voices you have permission to clone.
Octop (Python, MIT, ~5,700 stars) — Tencent Cloud's self-hosted AI assistant, and the pitch is the part most self-hosted tools skip: many users on one installation, with JWT isolation, tool approval, shell guardrails and PII redaction built in rather than bolted on later. You get a web console, a CLI and integrations into Feishu, DingTalk, QQ, WeChat, WeCom, Telegram and Discord; capabilities extend through connectors (OAuth plus MCP), a library of shareable expert role setups, RAG over your own documents, and IDE delegation to OpenCode or Claude Code behind permission gates. A beta coordinator schedules several experts across multi-step work, and an MIT licence with 48 contributors is a good signal for a July project. Version 1.0.2b4 landed yesterday with commits landing daily — this is a fast-moving beta, not something to hand your whole household tonight.
Codex Remote (Swift, macOS) — A menu bar app that puts Codex or Claude Code on a cloud machine you control, launched on Product Hunt yesterday. You give it a provider token, it provisions the server with OpenTofu across Hetzner, AWS, DigitalOcean, Linode, Vultr or Scaleway, installs the agent, and wires it up the way each one expects: Claude Code dials out to your account so it appears on claude.ai and on your phone, while the Codex app connects over SSH. The design decision worth copying is where your credentials live — on your Mac, with no hosted control plane in the middle — and MCP machine management is split so inspecting a machine and destroying one are separate opt-ins. It is days old and tiny, so watch it rather than standardize on it.
Offensive Security AI Models (curated list, ~280 stars) — A reference sheet of open-weight models fine-tuned for security work, and the table is the whole value: 27 models with parameter counts, context lengths, VRAM at Q4 quantisation, the uncensoring method and the licence — from a 7B model that fits in about 6 GB up to a 753B variant needing eight H200s. It separates security fine-tunes from general abliterated models, and the glossary actually explains what abliteration does to a model's weights, which most lists of this kind skip. Published September 2026 from model cards and vendor publications, so treat the benchmark figures as vendor-reported. Useful for authorised red-team and research work; not a shopping list for anything else.
Worth watching this week: whether VoiceStudio's model licences turn out to be as permissive as its code — that, not the feature list, decides whether teams can actually ship with it.
What is the last paid AI subscription you replaced with a self-hosted project — and did it stick? Tell us in the comments.
Sources: VoiceStudio (GitHub) · Octop (GitHub) · Codex Remote (GitHub) · Offensive Security AI Models (GitHub) · VoiceStudio benchmarks
