Dev open-sources an approval-queue tool that stops AI agents from sending anything until you sign off
paulofilip3 · reddit · 2026-10-03
Tired of agents ghostwriting emails and commit messages in obvious AI voice, a developer built and open-sourced a queue tool: agents can't send anything directly — everything lands as a draft with full context (what it is, who reads it, why the agent wrote it), and nothing goes out until the human edits and approves. Rejecting stops the agent.
Key details:
- In Claude Code it actually blocks sends (git commits, gh, Gmail, browser typing) until the exact text is approved
- One installer sets up Codex, Gemini CLI, Copilot and Cursor too, but only Claude Code has been really tested
- Emails and commits open in separate boxes (subject/body, title/description) with rich/preview/source modes
- Claude.ai and ChatGPT work via a connector, but it can only ask the agent to submit — not force it
- Can be paused globally or per folder (e.g. internal repos nobody reads)
More from coding & agent
- ChapterPal dev: frontier vision models consistently fail to spot obvious webpage conversion artifacts — burkov · 2026-10-03
- Dev finds Argon enough for nearly all coding tasks, misses it after switching to Opus 5.5 — m2saxon · 2026-10-03
- Air-gap file transfer via animated QR codes flashing at 10-30 frames per second — Thionne_WTZ · 2026-10-03
- Rust no longer the de facto best language for agents, dev argues — zack_overflow · 2026-10-03
- Knowledge search beats data connectors for agents: up to 88% faster, 95% cheaper — williamLberman · 2026-10-03
- INT21: 2 engineers direct AI to build 20 inference engines in 2 weeks, up to 2.4× faster than SGLang — bingxu_ · 2026-10-03