ai-automation-tools · open source · MIT

Scheduled automations that make things.

Claude Code automations that research, write, build, draw and compose, packaged as templates you can import into Cronsole. Run one by hand to see what it makes, or schedule it and let a collection build up.

Browse sample runs GitHub Quick start

48 recipes
7 groups
13 with a real run so far
Python 3.10+
Cronsole Registry v1

The recipes

One recipe, one automation

A recipe is a prompt, the skills it loads, a small Python runner and a Cronsole template. When it runs is your call: templates ship with a default schedule you change on import.

Apps & Games 9

A Manifest V3 extension, a playable HTML game, a 3D-print spec.

Writing 8

Short stories, a serialized novella, alternate histories.

Food & Maps 4

Safety-checked recipes with a Cook's Card, this week's viral dish, interactive maps.

Research Digests 12

Agent tooling, trending repos, a model benchmark watch, a nine-topic news digest.

Images 9

Comics, posters, 4K wallpapers, SVG icon packs.

Audio 4

Songs, a two-step debate podcast, narrated shorts.

Repo Ops 2

A review of every open pull request in one repo, and one small additive change to a static web app. Both default to a report and touch nothing else.

13 of the 48 recipes have had a real run so far; the rest have only run with the engines faked. The roadmap tracks which.

How it works

A complete folder or nothing

The agent works in a private staging folder. The runner checks every file it promised, renders the media, and only then moves the folder into place.

run.py ─▶ core/runner.py ─▶ claude --print, inside a staging folder
               │                └─ writes the files, returns JSON (title, summary, prompts)
               ├─ checks the files, removes strays, runs the recipe's own checks
               ├─ renders images, speech and music from what the agent described
               └─ moves the folder into place, writes artifact.json

Boxed in

The agent can only write to its staging folder. Claude runs under --permission-mode dontAsk with an explicit tool allowlist, and anything else is denied.

No keys in the agent

The agent never holds a media key. It describes the cover or writes the composition plan, and the runner does the rendering.

Same run everywhere

Skills ship in the repo and are copied into each run, so a run behaves the same on every machine.

Any agent CLI

Claude Code is the default. Codex, Gemini CLI, OpenCode, Kimi Code and Grok Build each run through one JSON profile in core/engines/clis/.

Then python scripts/build_viewer.py --output-dir ~/ai-output gives you one page for everything the recipes made: collections, covers, search, rendered Markdown, playable games and apps, galleries and media players. It opens from disk; no server needed. See it with sample runs.

Quick start

Run a recipe

You need Python 3.10+ and Claude Code signed in, with either a Claude subscription or an ANTHROPIC_API_KEY.

bash

$ git clone https://github.com/ai-automation-tools/ai-automation-recipes
$ cd ai-automation-recipes
$ pip install -r requirements.txt
$ cp .env.example .env          # add your image key and RECIPES_OUTPUT_DIR
$ python recipes/app-idea-card/run.py --output-dir ~/ai-output --dry-run   # print the prompt, no agent call
$ python recipes/app-idea-card/run.py --output-dir ~/ai-output             # run it for real

Output	Key	Without it
Covers	`GEMINI_API_KEY` or `OPENAI_API_KEY`	Published without a cover
Image recipes	`GEMINI_API_KEY` or `OPENAI_API_KEY`	The run stops before the agent starts
Speech	`GEMINI_API_KEY` or `ELEVENLABS_API_KEY`	The run stops before the agent starts
Music	`ELEVENLABS_API_KEY`	The run stops before the agent starts

ffmpeg is optional: with it, audio is MP3 and voiceovers get a vertical video. Voices are always premade presets; nothing here clones a voice. Claude usage counts against your subscription or API account, and each image is one billed image call.

Schedule them

Import into Cronsole

Cronsole imports the recipes as templates, in two editions. Take everything with the ai-recipes pack, one group pack, or single templates. A template never takes a key as a parameter.

Cronsole edition

Needs Python and the agent CLI where Cronsole runs, nothing else: the first run fetches a pinned, sha256-checked runtime into your cache. Keys are optional task secrets. Best for importing and forgetting.

recipe-<name>-cronsole

Clone edition

Needs a clone of this repo and a .env. Each command is a plain command line that cron, Task Scheduler or launchd can run as it is. Best for editing prompts, or any scheduler.

recipe-<name>

The Cronsole edition is a pilot covering short-story, comics and podcast-episode. The registry README has the details.

Resources

Links

Sample runsThe viewer, filled with real recipe output Source on GitHubai-automation-tools/ai-automation-recipes Recipe indexAll 48, with what each makes and needs Docs homeThe map of every document ArchitectureThe run loop, artifact.json, engines and skills Adding a recipeThe four files and the checks before a PR Cronsole registryTemplates and packs RoadmapLive-test status and what's left ChangelogEvery change, newest first CronsoleImports these recipes as templates automated-dev-pipeline skillFor scheduling a whole pipeline, in agent-skills IssuesReport a bug or request a feature