Adding Manual Approval Gates to AI Agents
sekyr95 · reddit · 2026-07-13
The author built an MCP server to add a structured gate for agents, ensuring actions "must be approved by a human before execution."
How it works
- The agent first calls impripushaction to submit the pending action
- The request enters an approval inbox with a status of 'pending'
- Humans can view, approve, reject, or even edit the text first
- The agent waits on impriawaitdecision and can only continue execution upon receiving 'approved'
Design motivation
The author emphasizes that this is not simply writing "please confirm first" in the system prompt, as models might bypass it. Instead, it makes approval part of the process structure, genuinely halting the agent before execution.
Other info
- Supports Slack / Discord / Telegram button approvals
- Includes an audit log recording who approved what
- Exposed via a REST API, and can also be used from MCP clients like Claude Code, Cursor, and Desktop
- The project is open-source under the MIT license, prioritizes self-hosting, and offers an early hosted version
The author admits the project is still in its early stages with few real users, primarily gathering feedback, especially on "where human intervention points should be placed."
Related event: New MCP Server Adds Human-in-the-Loop Approval for AI Agents(2 posts)→
More from coding & agent
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26