Claude Code desk stack gates writes with jev-guard scoring
A developer chains Claude Code, Workspace CLI, Playwright MCP, graphify and jev-guard into a six-step workflow where every write is scored 0–3, denied at 2.5 and asked at 1.5, at roughly $0.04 per 1,000 calls.
A developer stacked 4 repos into one desk. Claude Code, 133.9k-star graph, guard at 4 cents per 1,000. Six steps. Each one feeds the next. → OPEN a folder called AGENT. One rule in CLAUDE.md: save every decision to memory/. Next session inherits the last. → WIRE the Workspace CLI (31.8k stars). Gmail threads, calendar, Drive. Text back, not screenshots. → MOUNT Playwright MCP (38.6k stars). Pages return as structured nodes the agent reads like a screen reader. → INDEX with graphify. Tree-sitter parses code locally, no LLM call. The folder becomes a graph of typed edges. 133.9k stars, YC S26. → GATE with jev-guard. Every write scored 0 to 3, deny at 2.5, ask at 1.5. Reads skip the check. About $0.04 per 1,000 calls (README estimate). → CLOCK /schedule. Every morning at 7:30: what arrived, what was drafted, what waits for a yes. Two middleware tools at 31.8k and 38.6k stars. Both return structured text, not screenshots. Every layer after them, the graph and the guard, runs on parsed data. Strip GATE and the next bad page sends your mail. Skip INDEX and every session rescans the folder. Kill CLOCK and the desk dies after one chat. The order is the product. Full call journal, same numbers, step by step in the terminal below ↓
This post quotes another post: x.com/0xkarmi/status/2114102976788808056