Skip to content

Systemswebnoffice/agentic-os

BranchesTags

Repository files navigation

Agentic OS

Turn Claude Code into your Agentic Operating System.

Agentic OS gives Claude Code personality, memory, and skills so it works like a team member, not a chatbot. It remembers your brand voice, learns your preferences over time, and runs proven methodologies instead of winging it every session.

Quickstart

git clone https://<YOUR-TOKEN>@github.com/simonc602/agentic-os.git
cd agentic-os
bash scripts/centre.sh

Test edit: Added this line to test file editing permissions at 2026-04-16.

Replace <YOUR-TOKEN> with the access token from the Agentic Academy classroom.

On first launch, centre.sh runs the guided bootstrap automatically. It checks your system, prepares the local files Agentic OS needs, repairs missing dependencies when needed, and asks the one-time setup questions.

When it finishes, open Claude Code. It automatically detects you're new and walks you through building your brand foundation -- voice, positioning, and ideal customer profile.

On Windows, use:

powershell -File scripts\centre.ps1

Launching the command centre

After the first guided launch, you can keep using:

centre

That's it. The centre command reuses the saved launcher state, repairs missing bootstrap files silently when needed, starts the Next.js dev server, and opens http://localhost:3000 in your browser.

install.sh and setup.sh still exist for manual maintenance:

  • bash scripts/install.sh runs the guided installer directly.
  • bash scripts/install.sh --repair repairs only the local bootstrap files.
  • bash scripts/setup.sh refreshes dependency checks without launching the UI.

On Windows, the guided install can optionally add centre to both Windows PowerShell and PowerShell 7 profiles. If you prefer not to install the shortcut, keep using powershell -File scripts\centre.ps1.

Compatibility note: Agentic OS remains Claude-first, but the shared project instructions now live in AGENTS.md. Claude reads them through CLAUDE.md, and Codex can read AGENTS.md directly.

What You Get

Agentic OS is built on three layers:

  1. Agent Identity -- Personality (SOUL.md), your profile (USER.md), and session memory. This is what makes it feel like working with someone who knows your business.

  2. Skills -- Modular capabilities that can be added or removed. Each skill follows a tested methodology and self-improves as you give feedback -- corrections go directly into the skill, not just a note.

  3. Brand Context -- Your voice, positioning, and ideal customer profile. Skills load only what they need, so output stays focused and on-brand.

Core Skills (always installed)

Skill What it does
meta-skill-creator Build custom skills for your business
meta-wrap-up End-of-session memory and learning capture
mkt-brand-voice Extract your brand voice from content or build it from scratch
mkt-positioning Find angles that make your offer stand out
mkt-icp Define your ideal customer so every skill speaks to them

Optional Skills

Skill What it does API key needed
tool-humanizer Strip AI patterns from any output --
tool-firecrawl-scraper Scrape JS-heavy websites FIRECRAWL_API_KEY
tool-youtube Pull YouTube transcripts and channel listings YOUTUBE_API_KEY (channel mode only)
str-trending-research Research trending topics across Reddit, X, and the web OPENAI_API_KEY + XAI_API_KEY
mkt-copywriting Sales copy with 7-dimension scoring --
mkt-content-repurposing Turn one piece of content into posts for 8 platforms --
mkt-ugc-scripts Short-form video scripts with hook library --
viz-excalidraw-diagram Architecture and workflow diagrams --
viz-nano-banana AI image generation in 5 visual styles GEMINI_API_KEY
viz-ugc-heygen AI avatar videos via HeyGen HEYGEN_API_KEY
str-ai-seo Optimize content for AI search engines and LLM citations --
viz-interface-design Design dashboards, admin panels, and SaaS UIs --
ops-cron Schedule recurring Claude Code tasks --
tool-stitch Fetch UI designs from Google Stitch projects gcloud auth
viz-stitch-design Design and iterate on UI screens with Google Stitch gcloud auth

GSD (Get Stuff Done)

GSD is a project management framework for Claude Code. It's a separate install (not bundled with Agentic OS) that adds structured planning, execution, and verification for complex multi-step projects.

Install GSD:

npx get-shit-done-cc@latest

Use it when you're building something with multiple phases -- a product launch, a new feature, a migration. It handles planning, execution, verification, and session continuity.

Key commands:

Command What it does
/gsd:new-project Start a new project with deep context gathering
/gsd:plan-phase Plan a phase with research, task breakdown, and verification
/gsd:execute-phase Execute a plan with atomic commits and state tracking
/gsd:progress Check where you are and what's next
/gsd:debug Systematic debugging with persistent state
/gsd:quick Quick task with GSD guarantees (commits, tracking)
/gsd:verify-work Validate features through conversational testing
/gsd:pause-work Save context for resuming later
/gsd:resume-work Pick up where you left off
/gsd:help See all available commands

GSD and Agentic OS complement each other. Agentic OS handles brand context and skill-driven content production. GSD handles structured project execution when you're building something with phases and milestones.

Managing Skills

bash scripts/list-skills.sh                  # See what's installed and available
bash scripts/add-skill.sh mkt-copywriting    # Add a skill
bash scripts/remove-skill.sh viz-nano-banana # Remove a skill

Dependencies are resolved automatically. If you add a skill that needs another skill, both get installed.

Updating

bash scripts/update.sh

This pulls the latest changes from upstream -- new skills, improved methodologies, bug fixes. Your brand context, memory, projects, and API keys are never overwritten.

If you've customised any skills (via feedback, Rules additions, or direct edits), the update script detects this and shows you a diff for each changed skill. You choose per skill: accept our upstream changes, or keep your version. Either way, your version is backed up.

If new skills are available, the script tells you what was added and how to install them.

If the update fails with an authentication error, the access token has been rotated. Grab the latest token from the Agentic Academy classroom and update your remote:

git remote set-url origin https://<NEW-TOKEN>@github.com/simonc602/agentic-os.git
bash scripts/update.sh

API Keys

Most skills work without any API keys. Some are enhanced with external services (web scraping, image generation, video creation). All keys go in your .env file.

To see every available key with descriptions and signup links:

cat .env.example

Skills will tell you when they could use a key you haven't added yet, and they always offer a fallback so your work isn't blocked.

Scheduled Jobs (Cron)

Run tasks automatically through a managed runtime shared by the Command Centre and the CLI daemon. Drop a markdown file into cron/jobs/, then choose which host should keep the scheduler alive.

How it works

  1. The same cron core discovers jobs from the root workspace and every clients/* workspace
  2. It evaluates schedules, catch-up windows, dedupe, retries, heartbeat, and leadership in one place
  3. Matching jobs enqueue tasks and execute them headlessly via claude -p
  4. Per-job status still lands in cron/status/, logs still land in cron/logs/, and runtime state lives in .command-centre/
  5. Only one runtime host becomes leader at a time, so the UI and daemon never double-fire the same minute

Choose the host

Command Centre UI

  • The scheduler runs in-process while the Command Centre server is running
  • It stops with the server
  • It always starts the queue watcher, but only schedules jobs if it wins the leader lock

CLI daemon

  • Use this when you want scheduling to continue while the UI is closed
  • Start and stop it manually
  • It writes PID, lock, heartbeat, and logs to .command-centre/

Manage the daemon

Mac/Linux

bash scripts/start-crons.sh
bash scripts/status-crons.sh
bash scripts/logs-crons.sh
bash scripts/stop-crons.sh

Windows (PowerShell)

powershell -NoProfile -ExecutionPolicy Bypass -File scripts\start-crons.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\status-crons.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\logs-crons.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\stop-crons.ps1

Compatibility note: install-crons and uninstall-crons still exist, but they are deprecated wrappers around start-crons and stop-crons. They no longer register anything with Task Scheduler, launchd, or crontab.

The managed runtime uses your Claude plan credits. Each run costs roughly $0.01-0.05 (haiku), $0.05-0.25 (sonnet), or $0.25-2.00 (opus) depending on the model and task complexity.

Create a job

Each job is a markdown file in cron/jobs/ with YAML frontmatter and a prompt body:

---
name: "My Daily Research"
time: "09:00"
days: "weekdays"
active: "true"
model: "sonnet"
# Optional
# notify: "on_finish"
# description: "Researches trending topics and saves a daily briefing"
# timeout: "30m"
# retry: "0"
---

You are running as a scheduled job for Agentic OS.

Read CLAUDE.md for system context.

Task: [Your task here]

Save output to: projects/[folder]/{today's date}_[name].md

Or just ask Claude: "schedule a job to [do something] every morning" -- the ops-cron skill handles the rest.

Schedule options

Setting Examples
Exact time time: "09:00" or time: "09:00,13:00,17:00"
Every N minutes time: "every_5m", "every_10m", "every_30m"
Every N hours time: "every_1h", "every_2h", "every_4h"
Days days: "daily", "weekdays", "weekends", "mon,wed,fri"
Model model: "haiku" (cheap), "sonnet" (default), "opus" (powerful)

Full reference: cron/templates/schedule-reference.md

Notifications & status

  • OS notifications -- jobs send a native notification when they finish. Control this with the notify field: "on_finish" (default, notifies on success and failure), "on_failure" (errors and timeouts only), or "silent" (never notify).
  • Smart silence -- monitoring jobs that find nothing to report can suppress their notification automatically. The job's prompt tells Claude to end with [SILENT] when there's nothing actionable. The job still logs normally -- you just don't get pinged for "all clear" results.
  • No duplicate runs -- if a job is still running when the next scheduled trigger fires, the new run is skipped. This prevents slow jobs from piling up. If a previous run crashed without cleaning up, the system detects the stale state and recovers automatically.
  • Status tracking -- each job writes its result to cron/status/. The managed runtime also tracks leadership, heartbeat, PID, and daemon logs in .command-centre/.
  • Catch-up on wake -- if your laptop was closed during a scheduled fixed-time job, it runs automatically when the machine wakes up. Interval jobs (every_Nh) resume on the next matching interval without catching up.
  • Timeout -- prevents runaway jobs. Default is 30 minutes. Configure per job with the timeout field (e.g., "5m", "1h", "90s"). If a job exceeds its timeout, the process is killed and the result is recorded as timeout.
  • Retry -- set retry: "1" (or higher) to automatically re-run a job on failure. Each retry gets the full timeout. Default is 0 (no retries).

Manage jobs

Action How
Pause a job Set active: "false" in the job file
Resume a job Set active: "true" in the job file
Run a job now bash scripts/run-job.sh {job-name} on macOS/Linux or powershell -NoProfile -ExecutionPolicy Bypass -File scripts\run-job.ps1 {job-name} on Windows
Check logs cat cron/logs/{job-name}.log
List all jobs ls cron/jobs/ or ask Claude "what's scheduled?"
Start the daemon bash scripts/start-crons.sh or powershell -NoProfile -ExecutionPolicy Bypass -File scripts\start-crons.ps1
Stop the daemon bash scripts/stop-crons.sh or powershell -NoProfile -ExecutionPolicy Bypass -File scripts\stop-crons.ps1
Check runtime status bash scripts/status-crons.sh or powershell -NoProfile -ExecutionPolicy Bypass -File scripts\status-crons.ps1
Show daemon logs bash scripts/logs-crons.sh or powershell -NoProfile -ExecutionPolicy Bypass -File scripts\logs-crons.ps1

Stopping the daemon only stops automatic scheduling. Your job files in cron/jobs/ are never deleted.

Important notes

  • No OS scheduler fallback -- automatic scheduling no longer depends on Task Scheduler, launchd, or crontab.
  • UI behavior -- if you rely on the Command Centre host, scheduling stops when the server stops.
  • CLI behavior -- the daemon is manual. Start it when you want background scheduling and stop it explicitly when you do not.
  • Leader lock -- if the UI and daemon coexist, only one of them schedules jobs. The other host stays passive.
  • Legacy scripts -- run-crons is deprecated and no longer performs scheduling.
  • Existing sessions: Jobs run as separate headless processes -- they don't interfere with any open Claude Code session.

Windows Notification Theme

Windows notification visuals and default copy live in scripts/windows-notify.config.json.

  • app controls the Windows app identity, display name, Start Menu shortcut name, generated asset cache version, optional attribution, default toast duration, and the default layout (compact or hero).
  • assets.logoPath should point to the toast logo image. assets.shortcutIconPath should point to the Windows shortcut icon and must be an .ico file.
  • The repo-tracked asset folder for Windows branding is scripts/assets/windows-notify/.
  • assets.heroPaths.* can point to optional background images for hero mode. If blank or missing, the helper generates the hero background and overlays the event-specific text on top.
  • assets.generatedLogo controls the fallback logo artwork when the configured toast logo image is missing.
  • assets.variants.* controls the gradient palette, accent colors, sound, and emoji for each notification variant.
  • copy.interactive.* and copy.cron.* define the default status, subject, detail, variant, duration, and layout for each Windows notification event.
  • Supported template placeholders are {project}, {seq}, {duration}, {rawMessage}, {jobName}, {timeout}, {exitCode}, and {catchUpSuffix}.

Preview the default compact toast with:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/test-windows-notify.ps1 -Variant success

Preview hero mode explicitly with:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/test-windows-notify.ps1 -Variant success -Layout hero

If you change generated colors, layout, or badge text, either bump app.assetVersion in scripts/windows-notify.config.json or delete the cached folder under %LOCALAPPDATA%\AgenticOS\notifications\ so Windows rebuilds the generated assets.

Multiple Clients

If you work with more than one client or brand, just tell Claude: "add a client called [name]". It creates the workspace, and offers to switch you into it.

You can also do it manually:

bash scripts/add-client.sh "Client Name"
cd clients/client-name
claude

Each client has its own brand context, memory, and output. Shared methodology now lives in AGENTS.md at the root, with CLAUDE.md importing it for Claude Code. Skills and scripts sync automatically when you run update.sh.

For the full setup guide, see docs/multi-client-guide.md. For how projects work (single tasks, planned projects, GSD), see docs/projects-guide.md. For a quick reference, see docs/cheat-sheet.md.

File Structure

├── context/
│   ├── SOUL.md            <- Agent personality and behaviour rules
│   ├── USER.md            <- Your preferences and working style
│   ├── learnings.md       <- Accumulated skill feedback (gets smarter over time)
│   └── memory/            <- Daily session logs (auto-links to active projects)
├── brand_context/         <- Your brand data (voice, positioning, ICP)
├── .claude/skills/        <- Installed skill packs
├── command-centre/        <- Local UI and runtime app (versioned with the repo)
├── cron/jobs/             <- Scheduled job definitions
├── projects/              <- All generated output
│   ├── mkt-copywriting/   <- Single task category folders (Level 1)
│   └── briefs/            <- Multi-deliverable projects (Level 2/3)
│       └── q2-launch/     <- Project folder with brief.md
│           └── .planning/ <- GSD artifacts (Level 3, per project)
├── scripts/               <- Install, update, manage skills, watchdog
├── AGENTS.md              <- Canonical shared instructions
└── CLAUDE.md              <- Claude wrapper that imports AGENTS.md

Quality of Life

A few things baked in to make the day-to-day smoother:

  • CC Notify -- native OS notifications (Mac & Windows) when Claude finishes a task, needs permission, or is waiting for input. No more checking back every 30 seconds.
  • Auto-download -- binary outputs (images, videos, PDFs) auto-copy to your Downloads folder.
  • Humanizer gate -- every skill that produces publishable text automatically strips AI writing patterns before saving.
  • Clickable file paths -- every saved file shows the full path so you can click to open it.
  • Graceful degradation -- no skill breaks because something is missing. No API key? Free fallback. No brand context? Solid generic output.

Your Data is Safe

These are yours and are never overwritten by updates:

  • brand_context/ -- your voice profile, positioning, ICP
  • context/ -- your memory, learnings, session history
  • projects/ -- everything the system generates for you
  • .env -- your API keys (gitignored, never leaves your machine)

Need Help?

Head to the Agentic Academy Skool community. Post your question and the team or another member will help you out.

Built by Simon Scrapes @ Agentic Academy

About

Agentic OS based on Claude Code

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages