devcard

A live, embeddable dev stats card — powered by your real AI-assisted coding activity, not your keystrokes.

↑ A real, live card. Point at the bar and it names the language under the pointer.

WakaTime and friends are built around time spent coding. devcard is built around a different question: what came out of the session. A hook records the edits your agent makes — or, for any other tool, each commit — and your card updates from there.

• Live, not batch. An edit reaches your Worker within about half a minute, and the rendered card is cached for five minutes — not "synced last night".
• Content-private by architecture. Code, file names, paths and project names never leave your machine; nothing the hooks send has a field that could hold one. What does leave is activity metadata — when, which language, how many lines and bytes — plus a repository count and a random installation id. Exactly what, and what the card shows →
• Embeds anywhere an <img> works, follows the viewer's light/dark theme, and speaks en/pt/es.

Get one

Python 3.11+, Node 22+, npm, git, and a free Cloudflare account. The installer checks all of them before it creates anything in your Cloudflare account.

git clone https://github.com/augbastos/devcard && cd devcard && python install.py

It creates your D1 database, deploys your Worker, installs the capture hook, and prints your embed. Your deployment's settings go to a gitignored config, never into a tracked file, and re-running it is safe (a second machine copies the token and URL instead). Two questions — three in git mode, which also asks where your repos live. By hand instead →

Works with your agent

Pick one capture mode per machine; running both would count the same lines twice.

	Mode	Granularity
Claude Code	claude	Live, per-edit — the card moves while you code
Anything that commits through git	git	Per-commit, real diff stats

git mode hooks git itself rather than the agent, so anything that commits through git is covered — Codex, Cursor, aider, Windsurf, a local model, your own hands — with no per-tool integration to maintain:

python hook/install_git_hook.py "C:/path/to/your projects"   # a repo, or a folder of repos

Embed it

<img src="https://<your-worker-url>/svg?user=<you>" alt="devcard" />

For your GitHub profile: create a repo named exactly like your username and paste that line into its README.

?layout=	Size
full (default)	480×tall	Avatar, streak, language bar, 16-week heatmap, pinned repos, stats, badges
wide	840×~430	The same across a README's full column
banner	480×72	One-line strip for forum sigs
half	480×152	Header, lines, language bar, stats
vertical	280×264	Narrow column for sidebars

?theme= takes default (follows the viewer's system theme), dark, light, gentle, cyberpunk or terminal. ?langs=all names every language in the legend instead of grouping the tail.

Hover, inside a GitHub README

The bar draws every language to exact proportion, so a real account ends in a tail thinner than a pixel. Point at a segment and it names that language and its share — slivers included.

A README renders the card through an <img>, where no pointer event reaches the SVG, so the card is served in pieces instead: one image per language, each carrying its own tooltip. How, and what GitHub's sanitizer allows →

How it works

flowchart LR
    subgraph M["your machine"]
        CC["Claude Code<br/>PostToolUse hook"] -->|"every Edit / Write / Bash"| DB[("local SQLite<br/>offline-first")]
        GT["git post-commit hook<br/>(Codex, Cursor, aider, hand-typed…)"] -->|"git diff --numstat"| DB
        DB -->|"throttled, detached"| SY["devcard_sync.py"]
    end
    SY -->|"POST /ingest<br/>anonymized, token-gated"| W["Cloudflare Worker + D1"]
    E["your embed<br/>README · site · anywhere"] -->|"GET /svg"| W
    W -->|"rendered SVG<br/>cached 5 min"| E

Loading

Events land locally first, so capture works offline and never blocks your session. Each event carries a random installation id and its local row id, so a retried batch, a second machine or a deleted local database never counts anything twice — or mistakes new work for a duplicate. Rollup tables mean one render reads about a hundred rows rather than one per event ever recorded, which is what keeps a card inside the free tier.

Built to be checked

• Tests run against the real thing. The Worker suite runs inside workerd against a local D1; the git suite builds throwaway repositories and runs real git — merges, worktrees, core.hooksPath, husky, paths with spaces.
• CI with no Cloudflare account or secret: the Python suites on Linux and Windows at 3.11 and 3.14, the Worker on Node 22 and 24 — a fork's pull request runs all of it.
• Supply chain: a locked npm tree installed without install scripts, every Action pinned to a commit SHA, Dependabot, dependency review, Gitleaks over the full history, and CodeQL for TypeScript, Python and the workflows themselves.

Going deeper

Hover in a README	GitHub's sanitizer, measured, and the split card
What the numbers mean	Counting rules, merge commits, two machines
Privacy and security	What leaves your machine, and what stops it
Known limitations	Where it under- and over-counts, and why
Making it yours	Themes, layouts, badges, pinned repos
Development and CI	Both suites, the checks, and one deploy trap
Manual setup	The wizard's steps, by hand
Event retention	Why nothing is deleted, and how to measure your own

---

Built by Augusto Bastos · MIT