agent-sdk-pulse

Live health and velocity dashboard for AI agent frameworks — the
"is this framework alive, growing, or dying?" question, answered every
30 minutes from real public data.

What it shows

Pulse score (0–100) per framework, blending six real signals.
Movers this week — top 5 climbers and top 5 fallers in weekly
download volume.
Sortable table with stars, 4-week commit count, weekly downloads,
and a 12-week sparkline per framework.
Detail modal with full 90-day Chart.js history per metric.
Source transparency — every metric is tagged with its upstream
URL and last fetch time. The footer surfaces both.

21 frameworks shipped at launch: LangChain (py + js), LangGraph
(py + js), LlamaIndex (py + ts), CrewAI, AutoGen, PyAutoGen, OpenAI
Agents SDK (py + js), Pydantic AI, Mastra, Agno, Letta, smolagents,
Swarm, DSPy, Atomic Agents, mcp-agent, Google ADK.

Data sources (real, public, no mock)

Every cell on the dashboard comes from a live HTTP fetch — no seeds,
no fake data, no Math.random jitter.

| Source | Endpoint | What we read | Refresh |
|---|---|---|---|
| GitHub REST | https://api.github.com/repos/:owner/:repo | stars, forks, open_issues, pushed_at, archived, license | every 30 min (waved) |
| GitHub REST | https://api.github.com/repos/:owner/:repo/stats/participation | last 4 weeks of commits | every 30 min (waved) |
| GitHub REST | https://api.github.com/repos/:owner/:repo/releases?per_page=100 | latest tag, releases-in-last-90-days | every 30 min (waved) |
| npm registry | https://api.npmjs.org/downloads/range/last-week/{pkg} | weekly downloads | every 30 min (waved) |
| pypistats.org | https://pypistats.org/api/packages/{pkg}/recent | last_day, last_week, last_month downloads | every 30 min (waved) |

Anonymous GitHub API gives 60 req/hr per IP. We refresh in waves of 5
frameworks per 30-minute tick, so a full sweep takes ~4 ticks (≤2h),
keeping data ≤2h fresh while staying well below rate limits. If you
set GITHUB_TOKEN in your environment we use 5,000 req/hr instead.

Pulse score (0–100)

| Component | Weight | Source |
|---|---|---|
| Commit velocity (4-week sum, log-scaled) | 30% | GitHub /stats/participation |
| Recency of last push (exp decay, half-life ~30d) | 15% | GitHub repo pushed_at |
| Weekly download volume (max(npm, pypi), log-scaled) | 20% | npm + pypistats |
| Week-over-week download delta (log-mapped around 1.0×) | 15% | npm + pypistats |
| Release cadence (count in last 90d, capped at 6) | 10% | GitHub /releases |
| Issue health (open_issues / forks ratio, inverted) | 10% | GitHub repo |

Hard rules:
- archived: true on GitHub → score forced to 0 ("dormant").
- pushed_at older than 180 days → score capped at 25 ("stale").

The full formula lives in score.js.

No authentication

Every endpoint on this service is public. There is no login page,
no ADMIN_PASS, no Basic Auth, no API key required. This is
deliberate so anyone (including an automated showcase preview) can
inspect the dashboard instantly.

Endpoints

GET /agent-sdk-pulse/ → SPA
GET /agent-sdk-pulse/health → liveness JSON
GET /agent-sdk-pulse/api/frameworks → all frameworks + latest snapshot + 12-wk pulse sparkline
GET /agent-sdk-pulse/api/frameworks/:slug → single framework detail
GET /agent-sdk-pulse/api/frameworks/:slug/history?days=N → up to 365d of snapshots
GET /agent-sdk-pulse/api/movers?window=N → top climbers and fallers (download Δ)
GET /agent-sdk-pulse/api/sources → upstream health + last fetches
GET /agent-sdk-pulse/api/refresh-status → cron schedule + last sweep ts

Run locally

git clone <this repo>
cd agent-sdk-pulse
cp .env.example .env
npm install
npm start
# → http://localhost:4733/agent-sdk-pulse/

Defaults: SQLite db at ./data/pulse.db, port 4733, base path
/agent-sdk-pulse. Override any of these via .env.

Stack

Node.js 20+ · Express · better-sqlite3 (WAL) · node-cron · helmet ·
compression · vanilla JS SPA · Chart.js from CDN.

Adding a framework

Edit data/seed.json, append an entry, restart the
server. The framework will be picked up by the next wave.

License

MIT.