Tech stack

At a glance

Layer	Choice	Why
Language	Node.js (ESM, .mjs)	Claude Code’s native ecosystem. Hooks are shell commands — node is universally present.
Runtime deps	Zero	Stdlib only: node:fs, node:child_process, node:path, node:readline, node:os. Optional @anthropic-ai/sdk for SDK-fallback mode. No bundler.
AI — preferred	claude -p headless	Rides user’s existing Claude Code auth. Zero additional billing.
AI — Tier 0 (filter)	Local weighted scoring	Free. Kills ~70% of triggers before any API call. Runs in Node, no subprocess.
AI — Tier 1 (routine)	Claude Haiku 4.5	Cheap (~$0.003/call cached). Handles all automated distillation.
AI — Tier 2 (deep cleanup)	Claude Sonnet 4.6	Available via --model flag; no automatic trigger yet. Planned for Haiku drift correction.
Sync	Git (scoped to .wyren/)	Zero infra. LAN + WAN. Free history. Pluggable behind WyrenSync.
Storage	Filesystem markdown	Human-readable, git-diffable, native Claude context format.
SessionStart hook	additionalContext injection	Only Claude Code surface that injects hidden system context at session init.
Stop hook	Watermark + detached spawn	Tracks turn count; spawns distiller detached after threshold (5 turns or 2 min idle). Never blocks.
UserPromptSubmit hook	Live sync delta injection	Pulls remote on every user turn; injects only sections added since last injection. 3 s budget.
Distribution	install.sh / install.ps1 one-liners	Clone to ~/.claude/wyren/, wire hooks in settings.json, register wyren CLI via npm install -g.
CLI	bin/wyren.mjs	init, status, distill, broadcast-skill, install, update, uninstall, doctor, log. Pure Node.
Docs site	Astro Starlight → GitHub Pages	Markdown-native, search, dark mode, Mermaid, Node-aligned.

Totals

• External services: Anthropic API (only if claude -p unavailable), git remote.
• No database. No server. No auth. No frontend.

Non-choices (what we explicitly do NOT use)

Thing	Why not
TypeScript	Zero build step. Node runs .mjs directly; stdlib types are sufficient.
Bundler (webpack, esbuild)	Nothing to bundle.
Database	Filesystem markdown fits. No query surface needed.
Custom server	Git is the sync layer.
Auth	Trust model = same team. Git remote access is the permission boundary.
Docker	Nothing to containerize.
Observability stack	.wyren/log + wyren status CLI.
Test framework beyond node:test	Stdlib is sufficient.
React / Vue / Svelte	Docs site uses Starlight; plugin has no UI.

Why no MCP server?

MCP servers are tool-invocable only — Claude calls them as tools during a conversation. They cannot inject system context at session initialization.

Wyren’s core value is automatic context at start, which only hooks can deliver. An MCP server is a viable post-ship addition — on-demand query of teammate transcripts via a wyren_search(query) tool — but is intentionally deferred. See Future.

Why Node, not Python?

Two reasons:

1. Claude Code is Node. Hooks are shell commands; Node is guaranteed present on any machine running Claude Code.
2. Zero deps possible. A Python distiller would need anthropic package + pip + venv. The Node distiller shells out to claude -p via built-in child_process — zero install steps on teammate machines.