"Auto-Approve On, Coffee Time": A Developer Builds an Approval Gateway for His Own Coding Agent

A developer who published a tool-approval gateway for Claude Code introduced it by describing the habit that produced it. The opening of that post is reproduced below, verbatim.

Sooner or later I auto-approve every tool call in my Claude Code session.

Still a strange feeling to basically give the AI Agent the full power of doing whatever on my PC.

So I started to build an AI Coding Tool Gateway that intercepts any tool call from my Claude Code Sessions and matches it with a defined list of rules. In an local dashboard I can see all activity logs and add safe commands to my whitelist. Additionally I started to add Token Reduction Logic that cuts away unnecessary context from tool responses -> Saving me some tokens. Auto-Approve On - Coffee Time.

- apichap, in a "Show HN" post on Hacker News, 24 September 2026. Quoted verbatim and unabridged; the original is at news.ycombinator.com/item?id=49830471.

What was posted

The project is a gateway that sits between a coding agent and the machine it is working on. Its repository describes three jobs: record every tool call the agent makes, trim noisy tool output - progress bars, install logs, passing tests - before the model reads it, and enforce one global allow-or-deny rule set, with anything unrecognised blocked until it is approved from a local dashboard. It starts in monitor mode, so nothing is blocked until the operator switches it to enforce.

The repository was created on 23 September 2026, is written in TypeScript under GPL-3.0, and is published to npm. The post itself drew little attention at the time of writing - two points and no comments - so read it as one developer's account, not as a trend line.

Why the first line is the interesting one

The product description is ordinary; developer tooling for permissions is not new. What is unusual is the admission that precedes it. The reason the gateway exists is not that the author wanted to review each action - it is that he had already stopped reviewing them, and wanted a way to keep working that did not depend on his own attention holding up.

That is a different failure mode from the usual caution about autonomous agents. It is not that a model does something unexpected; it is that the human in the loop wears down, and the permission prompt turns into a keystroke. Anyone who has run an agent for a long session will recognise the shape of it.

The responses people build to that wear-down are worth watching, because they set the norms the next wave of tooling inherits: a rule list that is checked by something other than a tired person.

What this is not

This is one post by one developer about software he wrote himself, and it is a launch announcement rather than a report from a customer. Treat the behaviour it describes as a signal that is legible, not as evidence of how widespread it is. We have not tested the tool.

Source and provenance

This article quotes publicly posted user-generated content. It is a quotation with attribution, not original reporting by TechQuire. Source: Hacker News (platform: hackernews), type: launch. Author: apichap. Posted: 24 September 2026, 13:43 UTC. Original: https://news.ycombinator.com/item?id=49830471. Linked project: github.com/dommar04/ai-coding-gateway. The record is indexed in our Community Signals pool; removal or correction requests go through our contact page.

Image: "Programmer writing code with Unit Tests" by Joonspoon, via Wikimedia Commons, CC BY-SA 4.0.