Open Source Radar — September 15: agents get eyes and a voice
Today's board is about what sits between an agent and everything else: the hookup to the live internet, the voice it speaks with, and the box it runs in when there is no network at all.
Agent-Reach (Python, 81,783 stars, MIT) — A capability layer that gives any command-running agent read and search access to the platforms where the useful information actually lives: X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, RSS. Each platform gets an ordered list of backends — a primary plus fallbacks — that the tool selects, installs, and probes with a built-in health check; when a route dies, it re-routes without user action (the maintainers cite Bilibili blocking yt-dlp this June as one such swap). Everything is free and open source, cookies stay on your machine, and the README now carries an unusual notice disclaiming any crypto token using the project's name — a measure of how far it spread. Roughly 82,000 stars since a February start, and the fix for anyone who has spent an afternoon fighting Reddit's 403 on a server IP.
VoxCPM (Python, 37,518 stars, Apache-2.0) — OpenBMB's tokenizer-free text-to-speech line, back on the trending board behind the VoxCPM2 release: a 2B-parameter model built on the MiniCPM-4 backbone and trained on over 2 million hours of multilingual speech. Instead of discretizing audio into tokens first, it generates continuous speech representations directly, which is where the naturalness claim comes from. It covers 30 languages plus nine Chinese dialects, designs an entirely new voice from a text description of gender, age, tone, and pace with no reference clip, clones a voice from a short sample with optional style steering, and outputs 48 kHz audio. The maintainers report a real-time factor around 0.3 on an RTX 4090, improving to about 0.13 with vLLM-based serving. Apache-2.0 weights, commercial use allowed — the serious open alternative for voice agents and narration when per-character API pricing or data residency kills a hosted option.
Project NOMAD (TypeScript, 37,077 stars, Apache-2.0) — An offline-first knowledge and education server: one installer on a Debian-based machine brings up a browser-accessible Command Center that orchestrates a stack of containerized tools — offline Wikipedia, medical references and thousands of ebooks through Kiwix, Khan Academy courses with progress tracking, downloadable regional maps, data tools, notes, and a one-click app catalog. The AI layer is the interesting part for this audience: local chat over Ollama (or any OpenAI-compatible backend like llama.cpp or LM Studio) with document upload and semantic search through a vector store, so the knowledge base stays queryable with no internet. It's aimed at fieldwork, disaster response, and schools without bandwidth, and it wants a GPU for the AI components. The most complete packaged version of the run-it-yourself wave this board has been tracking all week.
oh-my-hermes (Python, 2,309 stars, MIT) — An all-in-one plugin for Hermes Agent, the open-source agent framework from Nous Research. It pitches itself as an operating layer rather than a replacement: problem framing, workflow selection, research and creation pipelines, coding handoffs, and project memory with explicit evidence boundaries, all running on top of the agent's native skills. Young — started in June — but moving fast, with commits landing today and a docs site covering each capability's claimed impact. Worth a look if you run an agent framework daily and care about the gap between what the model did and what it can show it did.
Worth watching this week: whether scraper-routing layers like Agent-Reach become standard agent infrastructure — or the first thing the platforms decide to block.
Which platform is the hardest to get your agent into right now? Tell us in the comments.
Sources: Agent-Reach (GitHub) · VoxCPM (GitHub) · VoxCPM2 weights (Hugging Face) · Project NOMAD (GitHub) · Project NOMAD site · oh-my-hermes (GitHub)
