Your GitHub stars, turned into a decision tool. Astrolab maps every repo you have starred into a 3D knowledge graph, scores each one's health and lifecycle, and writes curated landscape reports that answer "which tool should I use for X?", using only repos you have already vetted.
→ Open the live app · no sign-up, no install
Currently mapping 2,408 starred repos · 7,966 similarity edges · 39 communities · 29 landscape reports · refreshed weekly.
Stars are write-only memory. After a thousand of them you can't answer basic questions: what did I actually save about RAG? which of these is dead? if I need a chart library today, which one do I already know about?
Astrolab answers them. It does all the work offline at build time, with no LLM in the hot path. You get a graph you can fly through, reports you can read, and a search box you can ask.
| You want to… | Astrolab gives you |
|---|---|
| Pick a library for a new project | A landscape report ranking the candidates per task, with advantages, drawbacks and "best for" |
| Avoid adopting a dying dependency | Lifecycle stage, health score and bus factor side by side in Risk |
| See what's gaining momentum | Week-over-week star deltas and the Trending Now report |
| Find what you're missing | A discovery pass that searches GitHub for repos that belong in a report but aren't in your stars |
| Feed your tools and agents | Plain JSON and Markdown outputs that other projects and skills build on (see below) |
Every star as a node, positioned by a force simulation that runs at build time, not in your browser. Color is community, size is stars + PageRank. Click to focus, drag to orbit.
A co-occurrence graph of tags across the corpus. Shows the shape of your interests rather than the individual repos.
Prewritten analytical queries over the dataset: where developers actually work (by 90-day author counts), the classic core of your stars, hot repos, bus-factor risks, declining projects worth replacing, and PageRank-central cluster leaders. Every view exports to JSON or CSV.
Lifecycle stage, health score, bus factor, and top-author concentration surfaced together, so single-maintainer and abandoned dependencies stand out before you adopt them.
Searchable and sortable across every metric in the dataset: stars, forks, language, lifecycle, health, activity.
Full metric comparison plus the shortest path between them through the graph — which shared topics, authors, or intermediaries connect two projects.
The largest feature. Each report is a deterministic Python generator over the local dataset — hand-curated taxonomies, master comparison tables, per-task rankings, graph analysis, and maintenance risk. No API calls at generation time, so they're fully reproducible and cost nothing to rebuild.
Covering RAG tooling, agent harnesses, MCP servers, document extraction, LLM evaluation, voice agents, fine-tuning stacks, charting libraries, trending momentum, and more.
Questions answered with the graph's community summary as grounding context, via any OpenAI-compatible endpoint. Optional — every other tab works without a key.
The dataset and reports are plain files (data/classified.json,
public/data/graph.json, reports/*.md), so other projects can build on them.
These are the ones that did.
- hoshi (private): a self-hosted graph-vector index answering "what existing
repos solve this problem?", with the license on every hit. It was seeded from
Astrolab's
classified.jsonand similarity graph and now indexes 1,873 repos. - hf-models-analyzer (in progress): the same engine pointed at Hugging Face. It answers "given this task, this VRAM budget and this license, what should I run?" across 1,131 models and 1,198 datasets. It reuses the build-time precompute, the Python report orchestrator and the hard-won GitHub-client lessons.
- Repo-picker skills (private): Claude Code skills that turn a landscape
report into a reusable "which tool for X" advisor. Each one routes a use case
into the report's curated prose, then adds live health data (Active / Stable /
Slowing / Stalled).
scripts/build_viz_skill.pyin this repo generates the chart-library picker and re-generates it after every data refresh.
- kogitsune: a launcher for lean Claude Code sessions. It came out of reading the Claude Code setups report, which showed that per-session skill selection was the gap no starred tool filled. The repo was started the same day (2026-06-15).
- Stack decisions in new projects: reports get copied into other projects as their tool-choice reference. For example, the NotebookLM stack report (33 tools in 8 categories) became the design doc for one project's ingestion, retrieval and TTS picks.
- Chart-library choice everywhere: the picker skill answers "what should I chart this with?" from the Charting stack report while working in any other repo.
- Keeping the stars themselves useful:
scripts/discover.mjssearches GitHub for repos that belong in a report but are missing from the stars, and approved ones get starred. One pass added 14. The weekly refresh warns about any curated repo that was renamed or deleted upstream.
The whole design principle: do the expensive work at build time, ship a static file.
GitHub API
↓
scripts/sample.mjs # list the stars
scripts/ingest.mjs # fetch full repo + author data
scripts/classify.mjs # lifecycle stage + health score (no LLM)
scripts/precompute.mjs # kNN graph + Louvain communities + PageRank
# + 300-tick force simulation → frozen positions
↓
public/data/
graph.json # ~1.8 MB — nodes/links ready for the browser
graph-context.json # compact community summary for LLM grounding
↓
scripts/reports/*.py # 29 generators → reports/*.md + meta sidecars
scripts/reports/build_index.py
# snapshots the vintage, regenerates every report,
# injects chart SVGs, writes public/reports/index.json
Browser (React + Three.js)
→ loads graph.json and renders immediately (zero graph math in the browser)
Vercel Function /api/ask
→ loads graph-context.json, calls any OpenAI-compatible endpoint,
behind a daily/weekly spend cap kept in Upstash Redis
Snapshots in data/snapshots/ (one per vintage) power the ▲/▼ star-trend deltas
and the momentum-based Trending Now report.
npm install
cp .env.example .env # add GITHUB_TOKEN
npm run refresh # fetch → classify → precompute
npm run dev # → http://localhost:5173The repo ships with committed data, so npm run dev works before you ever run
refresh — you'll just be looking at someone else's stars.
npm run refresh # full chain: sample → ingest → classify → precompute
npm run precompute # rebuild graph.json from existing classified.json
npm run reports # regenerate all 29 reports + indexrefresh needs GITHUB_TOKEN in .env (a classic PAT with no scopes is enough —
it reads public data; the token exists for the 5,000/hr rate limit). Everything
downstream of ingest runs entirely offline.
| Script | Input | Output |
|---|---|---|
scripts/sample.mjs |
GitHub username | src/data/sample-all.json (star list) |
scripts/ingest.mjs |
sample + --max-age |
src/data/raw-all.json (full repo data) |
scripts/classify.mjs |
raw JSON | classified-all.json (lifecycle + health) |
scripts/precompute.mjs |
classified JSON | graph.json + graph-context.json |
scripts/reports/build_index.py |
classified + graph | reports/*.md + public/reports/index.json |
A GitHub Action (.github/workflows/refresh-data.yml) runs the whole chain weekly
and pushes the result; .github/workflows/deploy.yml then ships it to Vercel. The
refresh needs one repo secret, GH_STARS_TOKEN (deploy secrets: see Deploy).
Reports are curated by hand, not generated by an LLM. scripts/reports/rag_tooling.py
is the canonical template:
TAXONOMY = {"owner/repo": ("Category", "one-line blurb"), …}
ADJACENT = [(name, "why it's excluded"), …]Then register the file in GENERATORS in build_index.py and run it. The generator
prints tools: N / N curated and warns on any curated name missing from the dataset —
that warning is how you catch upstream repo renames.
scripts/reports/render_html.py renders any report as a standalone, fully inlined
HTML page (CSS and chart SVGs embedded, works offline).
Provider-agnostic: the Vercel Function (api/ask.mjs) talks to any
OpenAI-compatible chat-completions endpoint. The live site runs Nebius Token
Factory with Qwen/Qwen3-235B-A22B-Instruct-2507.
| Var | Required | Default |
|---|---|---|
LLM_API_KEY |
yes | — |
LLM_BASE_URL |
no | https://api.z.ai/api/coding/paas/v4 |
LLM_MODEL |
no | glm-4.6 |
Spend cap. Every request reserves its estimated cost against a daily and a weekly budget held in Redis, so the cap holds across serverless instances. With a cap set but no reachable store, requests are refused (fails closed).
| Var | Default | Purpose |
|---|---|---|
BUDGET_DAILY_USD |
1 |
daily cap (0 disables) |
BUDGET_WEEKLY_USD |
5 |
weekly cap (0 disables) |
LLM_PRICE_IN_PER_M / _OUT_PER_M |
1 / 3 |
USD per million tokens — keep in step with the model |
KV_REST_API_URL + KV_REST_API_TOKEN |
— | Upstash Redis (or the UPSTASH_REDIS_REST_* names) |
On Vercel: vercel env add <NAME> production, then redeploy — env changes only
take effect on the next deploy. Locally, drop them into .env. Legacy ZAI_*
names still work as fallbacks.
Provider recipes
| Provider | LLM_BASE_URL |
Example LLM_MODEL |
|---|---|---|
| Z.AI Coding Plan | https://api.z.ai/api/coding/paas/v4 |
glm-4.6 |
| Z.AI pay-as-you-go | https://api.z.ai/api/paas/v4 |
glm-4.6 |
| OpenAI | https://api.openai.com/v1 |
gpt-4o-mini |
| Groq | https://api.groq.com/openai/v1 |
llama-3.3-70b-versatile |
| Together | https://api.together.xyz/v1 |
meta-llama/Llama-3.3-70B-Instruct-Turbo |
| DeepSeek | https://api.deepseek.com/v1 |
deepseek-chat |
| Mistral | https://api.mistral.ai/v1 |
mistral-small-latest |
| OpenRouter (→ anything) | https://openrouter.ai/api/v1 |
anthropic/claude-sonnet-4-5 |
| Fireworks | https://api.fireworks.ai/inference/v1 |
accounts/fireworks/models/llama-v3p3-70b-instruct |
| Local Ollama | http://localhost:11434/v1 |
llama3.2 |
Without LLM_API_KEY the tab shows a configuration notice; everything else works
normally. Quota and billing errors are surfaced as a friendly notice, not a stack trace.
Hosted on Vercel; vercel.json holds the build command, rewrites, cache headers
and function config.
vercel link # once
vercel deploy --prod --yes # manual production deploy.github/workflows/deploy.yml deploys on every push to main and after each
weekly refresh. It needs three repo secrets: VERCEL_TOKEN, VERCEL_ORG_ID and
VERCEL_PROJECT_ID (the last two from .vercel/project.json).
graph.json and the reports are committed, so the build serves them without
running the pipeline.
| Layer | Libraries |
|---|---|
| UI | React 18, Tailwind CSS, Lucide |
| 3D graph | react-force-graph-3d, Three.js, d3-force-3d |
| Graph algorithms | graphology, graphology-communities-louvain, graphology-metrics |
| Data pipeline | Node.js ESM scripts, GitHub REST API |
| Reports | Python 3.12 (stdlib only — no dependencies) |
| LLM | Any OpenAI-compatible API |
| Hosting | Vercel (static + Functions), Upstash Redis for the spend cap |
scripts/
sample.mjs list stars for a user
ingest.mjs fetch full repo data from GitHub
classify.mjs lifecycle + health scoring
precompute.mjs graph build, simulation, LLM context
reports/
lib.py shared formatting + snapshot/trend helpers
build_index.py orchestrator: snapshot → generate → charts → index
render_html.py markdown report → standalone HTML page
*.py 29 report generators (one per landscape)
data/
classified.json classified dataset (committed)
snapshots/ one per vintage — powers trend deltas
public/data/
graph.json pre-computed graph (committed)
graph-context.json LLM grounding context (committed)
src/lab/
GraphProvider.jsx loads graph.json, provides nodes/links via context
MapView.jsx 3D force graph
TopicMap.jsx topic co-occurrence graph
InsightFeed.jsx named analytical queries
AllRepos.jsx searchable repo list
Comparator.jsx side-by-side comparison + path finding
ReportsView.jsx landscape report reader
LabApp.jsx tab routing + Ask AI UI
api/
ask.mjs POST /api/ask → any OpenAI-compatible endpoint
scripts/lib/
spend-budget.mjs daily/weekly spend cap backed by Upstash Redis
MIT © Martin Kaiser