GWS OS — Learning-first Google Workspace OS built on gws CLI

[cheeky-amit]

What I built

GWS OS — an open-source Claude Code skill folder that turns gws into a multi-account Google Workspace operating system with memory that learns your patterns over time.

One command (/gws morning) scans all your accounts, prioritizes what matters, and gets smarter every session.

Why gws

I evaluated several options (gcloud, custom Python scripts, MCP proxy servers) before landing on gws. The killer feature for AI-agent workflows: gws exposes the raw Google API with structured JSON output. The AI constructs the --params JSON directly and parses the response. Thinnest possible layer between "I need this data" and "here it is."

After 4+ weeks in production (500-800+ invocations, zero auth failures), I built an entire orchestration system on top of it.

How it works

/gws morning    — full morning brief across all accounts
/gws triage     — categorize and act on unread email
/gws reply      — context-aware reply from the right account
/gws plan       — schedule intelligence (flags conflicts beyond time overlaps)
/gws brief      — pre-meeting attendee context
/gws learn      — see what the system knows, adjust trust, undo actions

Multi-account via GOOGLE_WORKSPACE_CLI_CONFIG_DIR env var per profile. Works cleanly.

Memory system follows the Clustered Associative Recall (CAR) Protocol — three-tier storage inspired by how human memory actually works:

• Tier 1: Raw episodes (action JSONL logs, 7-30 days)
• Tier 2: Compressed summaries (contact + topic markdown nodes)
• Tier 3: Schemas and patterns (weighted graph edges, permanent)

Trust progression: observe → suggest → assist → automate. Earned per action type per contact. The system never sends without confirmation.

Tech stack

• Transport: gws CLI (this project!) — all Google API access
• Intelligence: Claude Code skill folder (11 skills with bash preambles + Claude instructions)
• Memory: Flat markdown + JSONL + Python graph engine (1053 lines, stdlib only)
• Tests: 82 pytest tests across 4 phases
• License: MIT

Numbers

• 82 tests, all passing
• 11 skill commands, all complete with working preambles
• 15 shared shell functions in lib/gws-common.sh
• 1053-line Python graph engine with 5 subcommands
• Zero external dependencies beyond Python stdlib + jq

Try it

git clone https://github.com/cheeky-amit/gws-os.git ~/Projects/gws-os
cd ~/Projects/gws-os
bash setup
/gws onboard
/gws triage

Requires: gws CLI (this project), jq, Claude Code.

Validation report

The repo includes a gws CLI validation report and a production case study from 4+ weeks of daily use.

Feature requests for gws

Based on building this, two things that would make gws even better for multi-agent workflows:

1. Named profiles (gws --profile work gmail ...) — the CONFIG_DIR env var works but a first-class flag would be cleaner
2. Compact JSON mode (--format json-compact or --json-lines) — for piping to jq without the keyring backend line

Happy to discuss either of these in separate issues if helpful.

---

Built with gws 0.16.0 on macOS. AI orchestration via Claude Code.