Open Source Radar — September 22: the self-hosted swap
Today's board is one argument told five ways: the paid per-seat product and the self-hosted substitute are now the same product, and one of them runs on hardware you already own.
AutoClip (Python, ~8,600 stars, MIT) — Long podcasts, interviews, lectures and stream replays go in; a scored outline, a topic timeline and short clips with generated titles come out. It finds highlights from the transcript rather than watching the video, works from an existing subtitle file or transcribes locally, and exports with presets for the vertical-video platforms, burned-in subtitles and title cards included. It ships three ways — a desktop app for Apple Silicon Macs and Windows, a self-hosted web interface, and a headless mode any MCP-aware client can call — and the model is your choice: hosted Qwen, Gemini, an OpenAI-compatible endpoint, or a local model through Ollama. The interface and release notes now cover eight languages; the majority of the documentation is still Chinese, which is the one thing to know before you fork it. The draw is arithmetic: per-minute clipping services against a GPU you already paid for.
ai-memory (Rust, ~7,950 stars, MIT) — A shared long-term memory server for coding agents, now covering twenty-plus harnesses from Claude Code and Codex to Cursor, Gemini CLI, OpenCode and Grok. Quit one agent mid-task, open a different one in the same directory on a different machine, and it starts with a real handoff: where you left off, what failed, what is still open. The design decision that matters is that the source of truth is a git-backed wiki of ordinary markdown files — greppable, editable by hand, restorable — with a rebuildable index on top, and the default path makes zero LLM calls, so capture, search and handoffs work with no API key at all. Version 2.4.0 landed yesterday, and the project has tripled since we first flagged it in August at 2,400 stars; multi-user auth, attribution and an audit log are built in rather than sold as a tier. Worth a look if your team's "shared memory" is currently one person's laptop.
Laya (Python, ~13,400 stars, Apache-2.0) — An independent, non-autoregressive decision engine in the Jev mould, and this one ships the weights. A single 420-million-parameter checkpoint answers up to fifty typed questions in one forward pass — roughly 33 milliseconds per question on a T4 — in three shapes: pick a label, place a value on an ordinal rubric, or return a calibrated true/false probability. The probabilities are trained with proper scoring rules, which is what makes the confidence gate usable: act automatically above 0.85, escalate below it. Presets cover model routing, prompt guardrails, moderation and support triage, and the numbers in the repo's own benchmark put it ahead of Jev on typed decisions with far better calibration, a multilingual checkpoint handles 51 languages, and a free-tier notebook fine-tunes it on your own data. Two caveats stated plainly: those benchmarks are the maintainer's, and the Jev figures are third-party published numbers compared, not measured head-to-head. The honest pitch is control — a System 1 layer you can run air-gapped for nothing.
Claude for Financial Services (Python, ~36,000 stars, Apache-2.0) — Anthropic's vertical packaging, and the interesting part is that none of it is code. Named agents for pitchbooks, earnings reviews, valuation, general-ledger reconciliation, month-end close and KYC screening sit on top of skills, slash commands and data connectors, with partner plugins contributed by LSEG and S&P Global. Everything is markdown and YAML, deployable either as a Cowork plugin or as a Managed Agent cookbook behind your own orchestration, and the repository's own linter keeps the bundled skills from drifting. The newest addition is the advisors plugin, launched September 14 with Orion's connector pulling live portfolio and CRM data into the conversation while respecting the firm's existing permissions — then spun out into its own repository days later, which tells you how fast this packaging layer is still moving. The most interesting open-source release of the month is a big company's job function, written down.
NiubiGEO (TypeScript, ~4,800 stars, Apache-2.0) — A self-hosted way to see how AI assistants describe your product. It puts the same question to provider APIs with and without your brand name, keeps the raw answers and the sources they cited, and separates domain recognition from unbranded discovery — so you find out which competitors get named and, more usefully, which pages the models pulled from. Knowing that AI visibility shifts for reasons other than your changes, it reports variance rather than a single score. On Product Hunt today, and honest about the split: the measurement tool is free and open, while human-run testing and promotion are paid services. Its own documentation carries the limits — it observes provider API responses, not consumer chat apps, and it does not track traditional search rank. The correct first step for anyone whose traffic is quietly moving into answer boxes.
Worth watching this week: whether the local decision engine and the local memory server stop being side projects and become the default layers under everyone's agent loop.
Which paid per-seat tool are you most likely to replace with a self-hosted project this quarter? Tell us in the comments.
Sources: AutoClip (GitHub) · ai-memory (GitHub) · ai-memory v2.4.0 release · Laya (GitHub) · Claude for Financial Services (GitHub) · NiubiGEO (GitHub)
