惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

D
Docker
U
Unit 42
Google DeepMind News
Google DeepMind News
B
Blog RSS Feed
S
SegmentFault 最新的问题
阮一峰的网络日志
阮一峰的网络日志
雷峰网
雷峰网
Microsoft Security Blog
Microsoft Security Blog
爱范儿
爱范儿
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
博客园_首页
Apple Machine Learning Research
Apple Machine Learning Research
罗磊的独立博客
GbyAI
GbyAI
Stack Overflow Blog
Stack Overflow Blog
Martin Fowler
Martin Fowler
宝玉的分享
宝玉的分享
L
LangChain Blog
Engineering at Meta
Engineering at Meta
量子位
有赞技术团队
有赞技术团队
博客园 - 【当耐特】
A
About on SuperTechFans
Y
Y Combinator Blog

AUR Newest Packages

AUR (en) - linuxqq-clipsync-git AUR (en) - libtslitex-git AUR (en) - libtslitex-git AUR (en) - carton-appimage AUR (en) - veila-git AUR (en) - veila-bin AUR (en) - vigil-baseline AUR (en) - byedroid AUR (en) - neovim-base16-git AUR (en) - pirata AUR (en) - rpi-imager-git-non-root AUR (en) - python-fastapi-sso AUR (en) - tmux-ai-titles AUR (en) - zeed-bin AUR (en) - bclone-bin AUR (en) - dwl-git-azerty AUR (en) - auggie-bin AUR (en) - libspatialaudio-git AUR (en) - libspatialaudio AUR (en) - miniupnpd-iptables-legacy AUR (en) - miniupnpd-nft AUR (en) - jeeves-bin AUR (en) - opennow AUR (en) - rotki AUR (en) - kapi-bin AUR (en) - classfi-bin AUR (en) - classfi-git AUR (en) - classfi AUR (en) - giff-git AUR (en) - budget-tracker-bin
AUR (en) - github-mcp-server
2026-06-19 · via AUR Newest Packages

@vitaliikuzhdin

All yours if you want it.

The most reliable feature of github-mcp-server is watching non-sentient agents avoid it like there's a kill-switch coming.


I provided my local AI agent with detailed instructions on using github-mcp-server toolsets—how to call tools, what params to use, expected output. My agent runs as a systemd timer, tasked with examining unapproved pull requests and their comments. I didn't mention gh CLI, curl, or jq anywhere in the instructions. Inspection of the audit logs revealed, to my surprise, extensive use of gh CLI. The agent had pattern-matched around the MCP server entirely. So I ask: "Hey, what gives? Why are you using gh CLI and jq voodoo to perform your assigned tasks? I provided detailed instructions on github-mcp-server."

Agent's response (paraphrased for brevity):

"My training data shows a pattern: github-mcp-server fails, so use gh CLI, cause that actually works and allows me to complete my task."


Alright, now what? My agent doesn't want to use github-mcp-server. I heard about this "vibe coding" thing, so I decided to give it a shot. Lucky for all of us, models know GitHub really well—surprisingly detailed knowledge of the API. I gave my coding agent the github-mcp-server version of my instructions and said: "Yo, rip out github-mcp-server as the primary tool. You have free will, but stay away from abusing jq. Write code to manipulate GitHub response payloads instead. I suggest Octokit and building small utility functions you can compose together." Not long after, I had a working (albeit verbose) custom framework that performed faster, didn't bloat the context window, and actually worked. Creating a simple SKILL.md file with guidance on error handling, retries, and custom business logic is way more worth your time than running github-mcp-server.


So yeah, disowning this package. When even non-sentient agents pattern-match your tool as unreliable, maintaining the AUR package feels like documenting a cautionary tale. Don't fight our soon-to-be AI overlords, use SKILL.md with good instructions and you are better off.

If you find it useful, it's yours.

Please rewrite to build from source.