Every team I work with pays for two things that never talk to each other: the tools where the work lives — Jira, Basecamp, Slack, Zendesk — and an AI subscription that lives in a chat window. The glue between them is a person, copying context into the chat and pasting answers back out. That glue work is boring, constant, and completely automatable.
Contents
- One command, five minutes
- Five patterns, thirty-five tools
- The rules every recipe follows
- Where it stands honestly
- Follow along — and name your tool
So I built the glue once, properly, and open-sourced it. ai-automations is a collection of 35 plug-and-play AI automations for the tools your team already uses. Each one installs from a single interactive script and is running in about five minutes.
One command, five minutes
Every automation is a self-contained recipe: clone the repo, run ./setup.sh, pick your tool, answer a few questions — your project IDs, your tokens, your conventions. The installer copies battle-tested files into your repository, sets your secrets, and prints the two or three remaining manual steps with your real values already filled in.
There is nothing to host. Recipes run on GitHub Actions, woken by webhooks through free-tier relays — zero idle cost — or entirely on your own box if you prefer (the Docker runtime is Ollama-ready for fully self-hosted AI). And they work with whatever AI you already pay for: Claude Code, Codex, Gemini CLI, or any raw LLM API.
Five patterns, thirty-five tools
Under the hood the 35 recipes are five repeatable patterns:
- Project agents (Jira, Basecamp, Linear, Trello, ClickUp, Notion, Asana, Monday, Airtable, Shortcut, Todoist, Azure DevOps, GitHub, GitLab…) — move a card to a column and the AI analyses it against your codebase, discusses in the comments, and opens a pull request.
- Triage agents (Zendesk, Intercom, Freshdesk, Help Scout, Front, Slack, PagerDuty, Sentry) — tickets and alerts arrive pre-researched, with grounded draft replies saved as internal notes for a human to send.
- Notify agents (Discord, Teams, Telegram, LINE, Google Chat, Mattermost, Rocket.Chat) — AI-written ship and incident notes, in plain language, where the team already reads.
- Deploy & build agents (Vercel, Netlify, Buildkite) — a failed deploy becomes a triage issue with the log excerpt and likely cause; recurrences comment instead of duplicating.
- The odd ones — ask your codebase questions from a Figma comment, a Miro sticky note, or a Confluence page review.
The rules every recipe follows
The part I care most about is not the coverage, it is the posture. Every recipe obeys the same rules:
- Humans stay in charge. The AI writes comments and opens pull requests. It never moves cards, merges code, closes tickets, or deletes anything.
- Webhooks are doorbells, not data. A webhook only wakes the automation — every run re-fetches the truth from the tool's API, so duplicate or missed events are harmless.
- State lives in git. Per-item state is JSON on an orphan branch. No database, fully inspectable, versioned for free.
- Every run is auditable. The prompt, the full transcript, and the result are uploaded as CI artifacts.
- Secrets stay secrets. Installers never write tokens into files — credentials go into secret stores only.
Where it stands honestly
The Basecamp recipe is the production-proven one — it is the pattern the rest were built from. The other 34 are built and smoke-tested but marked beta: each carries a live-fire checklist in its README that I am working through, one tool at a time. If you hit an edge, open an issue — that is exactly the feedback the beta label is asking for.
Follow along — and name your tool
I am releasing these publicly as a series on LinkedIn: one tool every few days, with the five-minute setup shown each time. And the series has one standing offer — if your tool is not on the list, tell me and I will build the automation for it, free. The recipe contract makes new tools genuinely quick to add, and a request tells me exactly which one matters to real teams.
Everything is MIT-licensed and free: github.com/akshaypatel-ai/ai-automations.
Key takeaways
- Five repeatable patterns cover all 35 tools: project agents, triage agents, notify agents, deploy and build agents, and a handful of odd ones.
- The AI comments and opens pull requests. It never moves cards, merges, closes or deletes.
- Webhooks are doorbells, not data — every run re-fetches the truth from the tool's API.
- State is JSON on an orphan git branch, so there is no database to run.
- Every run uploads its prompt, transcript and result as CI artifacts.
- Installers never write tokens into files.
- One recipe is production-proven; the rest are beta and say so.
FAQ
Do I have to host anything?
No. Recipes run on GitHub Actions, woken by webhooks through free-tier relays, so there is no idle cost. You can also run them entirely on your own box — the Docker runtime is Ollama-ready for fully self-hosted AI.
Which AI does it use?
Whichever you already pay for: Claude Code, Codex, Gemini CLI, or any raw LLM API.
How much of this is production-proven?
One recipe. Basecamp is the pattern the rest were built from and the one run on real projects; the other 34 are built and smoke-tested but marked beta, each carrying a live-fire checklist in its README.
My tool is not on the list.
Name it and I will build the automation for it, free. The recipe contract makes new tools quick to add, and a request tells me which ones matter to real teams.
Working on something similar?
If you are wiring AI into the tools your team already uses, I am happy to look at it with you. Thirty minutes, no deck — bring the workflow that costs your team the most time and we will work out whether it is automatable. Get in touch.