juandi-net/alpha-oss

Alpha is a stream of alpha: it pulls from many sources (what you save: X bookmarks, GitHub stars, Hugging Face likes;
what it finds for you: your X Following and For You timelines, the Substacks you subscribe to, what the Hugging Face
accounts you follow release, what is trending on GitHub and Hugging Face, Substack search, papers that ship code,
Hugging Face datasets), keeps only the real signal, and turns it into actionable skills your AI agents use.
How: one hard judge for everything Alpha finds (your bookmarks teach it your taste; your projects' focus briefs tell
it what matters now), one warehouse, a feed you scroll instead of X, skills built and tested from what lands there,
and agents reading it all through the MCP server ("check Alpha for ...").
The rule throughout: the idea first, the vessel second. A post, repo, model, dataset, or paper carries a concept,
technique, finding, or capability. The judge weighs that idea, and suggestions put it to work with what your
projects already have. The artifact itself (installing the repo, running the model, loading the dataset) counts only
when it squarely meets a current need and is clearly worth the work.
## Quick start
macOS, Python 3.14, uv.
sh
git clone https://github.com/juandi-net/alpha-oss.git alpha && cd alpha
uv sync
uv run alpha setup # checks tools, creates .env and config.toml, says what is missing
# fill in .env (keys) and config.toml (projects, sources, context feeds)
uv run alpha run # first run: collect, discover, process
uv run alpha install # launchd: a run every 6 hours (each source keeps its own cadence) and the dashboard
open http://localhost:3460
You need an Anthropic API key (ALPHACLAUDEAPIKEY) and a TypeSafe key (TYPESAFEAPIKEY, for Jev). Every
source and context feed is off until you enable it in config.toml. alpha status shows what is set up.
## How it works
sources ──> (discovered only: judge) ──> warehouse (SQLite) ──> outputs: skills, project matches, dashboard, vault pages
context: focus briefs (what you work on)
Curated sources bring items you picked, such as your X bookmarks. They are trusted and stored as they are.
Discovered sources bring items Alpha finds, such as Substack posts. Every candidate goes through the shared
judge (judge.py): stage 1, Jev scores title and preview against your focus briefs; stage 2, the full text goes
to Jev and to Claude ([judge] model, Haiku by default), and it becomes a find when both call it "sauce"
(specific, backed by the author's own results, usable now; for a dataset or a model or Space: it presents an idea
you can put to work, in how its tasks or labels are designed, how it was built or evaluated, or a capability now
cheap or open enough to change what you can do, backed by a real card; you need not run it) and it repeats nothing
you already have (jev.py holds the definitions and IDEARULE).
When only one judge says sauce (Jev is strict about "usable now"; Claude often says yes to what you could act on
later), a new candidate becomes an idea: an item with a quiet Idea mark in the Feed and the Ideas filter. Rate
an idea α and it becomes a find; until then it feeds no skill.
Taste: Claude writes a short profile of what you keep and what you would call slop from everything you picked