1,244 GitHub stars turned into agent-ready memory. One Rust binary. Offline-first. Your coding agent gets facts plus judgment in one bounded object.
cargo install jev-stars| Check | Result | Rerun |
|---|---|---|
| Sync | 1,244 repos, 8 archived | jst sync + jst status |
| Search latency | 47-49ms on 1,244 rows | time jst search "rust" --limit 5 |
| Tests | cargo test green, 5 tests | cargo test |
| Curate ledger | rerun on done rows = 0 Jev calls | jst curate --limit 50 twice |
| Jev confidence | 0.72-0.79 with candidates, 0.37-0.46 without | see HANDOFF |
| Repo | Stars |
|---|---|
| sindresorhus/awesome | 512,141 |
| public-apis/public-apis | 484,174 |
| donnemartin/system-design-primer | 372,425 |
| vinta/awesome-python | 323,926 |
| obra/superpowers | 292,613 |
| practical-tutorials/project-based-learning | 285,170 |
jst syncPull stars via gh api, upsert SQLite. Needs net.jst search <q>FTS + --lang / --topic / --alive. Offline, ~47ms.jst context <q>Bounded agent JSON: health + Jev + evidence. Start here.jst statusDead/alive/license rollup. Offline.jst curateJev Choice + Score batch, hash-cached. Optional, needs key.jst exportSTARS.md awesome-list, archived skipped. Offline.jst mcpstdio MCP server: search, context, status, sync, export.$ jst context "rust mcp" --limit 1 repo: 1jehuang/jcode (20,200 stars, MIT) maintenance: active, pushed 2026-09-29 why: FTS(rust mcp) excerpt: The most RAM efficient harness
{
"mcpServers": {
"jev-stars": { "command": "jev-stars", "args": ["mcp"] }
}
}
5 tools: search, context, status, sync, export. Start with context.
cargo install jev-stars jst sync jst search "rust mcp"
curl -LO https://github.com/AkashPriyadarshii/jev-stars/releases/latest/download/jev-stars-windows.exe jst sync
git clone https://github.com/AkashPriyadarshii/jev-stars.git cd jev-stars cargo build --release ./target/release/jev-stars sync
Does jev-stars need an API key? No. Search, rank, context, status, and export work offline with zero keys. Only curate needs TYPESAFE_API_KEY, once, and results cache by content hash.
How is this different from GitHub's stars page? GitHub gives you a paginated list. jev-stars gives your agent a bounded context object: health, license, Jev score, and evidence excerpt in ~2k tokens.
Why not vectors in v0.1? FTS5 answers 1,244 rows in 47ms. sqlite-vec lands in v0.2 only when FTS measurably falls short.
No vectors in v0.1 (sqlite-vec lands in v0.2). No TUI, no dashboard, no release tracker. No star/unstar. Read-only memory over your stars.