yolo-labz

AI-native automation tooling for high-leverage operators — built compliance-grade: build provenance, signed attestations, and hard rate/cost ceilings are first-class citizens, not bolted on after the demo.

Plugin family for Claude Code, system daemons for macOS / Linux, and infra recipes that compose. Apache-2.0 + MIT throughout. SLSA L2 + Sigstore + reproducible builds.

Plugins (Claude Code)

Plugin	Stack	Purpose
wa	Go · whatsmeow · SQLite · JSON-RPC	WhatsApp daemon with append-only safety, allowlist + rate limiter + warmup ramp. Built so an LLM can use it without becoming a vector for spam or impersonation.
claude-mac-chrome	Bash · AppleScript · TypeScript	Multi-profile Chrome control on macOS. Reads Chrome's Local State catalog + drives React/Ember SPAs via cliclick at computed screen coords. Zero user configuration.
linkedin-chrome-copilot	TypeScript · Anthropic SDK	LinkedIn copilot: five user-stories (resume, draft-reply, book-slot, CV tailor, guardrails). Per-channel tone + transport + validation. PII-scanner CI gate.
kokoro-speakd	Python · ONNX	Persistent Kokoro TTS daemon. Loads model once, serves over unix socket. Sub-200ms warm-call latency. PyPI-published with PEP 740 attestations.
claude-classroom-submit	Python · Google Classroom API	Bypasses Drive Picker iframe via REST API + OAuth 2.0. End-to-end submission < 5s.
fand	Rust · launchd	Apple Silicon thermal daemon. Temperature-driven curves, SIGHUP reload, exact-pin reproducible builds.
anthropic-throttle-proxy	Python · HTMX	Self-hosted throttle proxy for the Anthropic API: AIMD rate ceiling for an agent fleet, live saturation dashboard, advisor loop. The spend/429 guardrail everything else runs behind.

Composition

Plugins compose. linkedin-chrome-copilot uses claude-mac-chrome for browser routing. Both sit alongside wa for cross-channel pipelines. kokoro-speakd narrates agent status without per-call model-load latency. fand keeps the laptop cool while all of the above run in parallel.

The agent loop becomes a graph of specialized capabilities, not a monolithic chat assistant.

Distribution + supply chain

gh attestation verify --owner yolo-labz: one command, no cosign install, confirms a release artifact was built by this org's CI from the tagged source. Every plugin's release pipeline produces:

• SLSA L2 build provenance via actions/attest-build-provenance (GitHub-native attestations; claude-mac-chrome adds SLSA L3 via slsa-github-generator)
• Dual SBOM: CycloneDX 1.7 + SPDX 2.3 (syft / anchore/sbom-action)
• Reproducible builds: SOURCE_DATE_EPOCH, -trimpath, -buildvcs=true where the toolchain supports it

Python plugins (kokoro-speakd, claude-classroom-submit) publish to PyPI via Trusted Publishing with PEP 740 attestations.

CI is hardened across every repo: SHA-pinned actions (40-char) with version comments, permissions: {} deny-all + per-job re-grant, and step-security/harden-runner egress auditing.

Tap

brew install yolo-labz/tap/ for CLI tools. Pre-notarized macOS binaries built from a Linux runner via rcodesign.

brew tap yolo-labz/tap
brew install yolo-labz/tap/wa

Constitution

Each repo carries a constitution.md (under .specify/memory/) with binding rules: hexagonal core, daemon owns state, no --force ever, CGO_ENABLED=0 where applicable, conventional commits, signed tags. Spec-driven development with citations: every "best practice" claim links to a primary source.

Author

Pedro Balbino · Senior SWE specializing in AI-native automation. Engineering writeups + plugin demos at blog.home301server.com.br · portfolio at portfolio.home301server.com.br · LinkedIn balbinopedro.