On this page
Every client you take on means rebuilding the same ops stack. The same audit checklist. The same break-even sheet. The same weekly report, in two languages. Foundry is that stack, already built, packaged so your AI coding agent can run it.
Marketing is engineering, so it should behave like software.
Foundry is a growth ops kernel for ad agencies. It is not a pile of prompts. It is a Python engine with deterministic commands, and the AI agent is the interface on top.
What you get on day one
Audit and diagnose. Say "audit my account" (or "راجع حسابي") and get a 0 to 100 score, a grade, the weakest dimension, the top fixes and a 30-day plan. Then ask what is wrong and get the single biggest constraint, with the findings ranked behind it.
Client reports in English and Arabic. Decision-first readouts written to be sent, not decoded, including a short WhatsApp digest.
Money maths that does not guess. Break-even CPA and ROAS. Real ROAS on cash-on-delivery accounts, judged on delivered revenue rather than platform-reported orders. Break-even cost per lead for lead-gen and WhatsApp funnels. Budget needed to exit the learning phase. VAT splits across MENA countries. The project describes its calculators as deterministic Python and says they print their assumptions alongside the figures.
No API access? Import a platform CSV export and the same diagnosis runs on it.
Campaigns, content, tracking, leads, creative, competitors, seasons and forecasts. Commands for campaign visibility, a content pipeline, tracking readiness, lead scoring, hook and angle ideas, competitive positioning, Ramadan, Eid and White Friday planning, and budget forecasts. 31 commands in all.
It keeps working between conversations
Six scheduled loops turn signals into work: Google Ads monitoring and budget-burn alerts, the same for Meta, a content engine that turns signals into briefs, a calendar that decides when and where they publish, a creative-brief step that writes image and video prompts, and a lead qualifier that scores inbound leads and drafts the outreach. The project describes maker/checker task contracts and a no-auto-publish policy.
One kernel, many clients
Each client lives in an isolated folder with its own credentials, strategy and history. The tenant loader checks exact folder names and rejects symlinked tenant paths and existing paths that resolve outside the client folder. Reports and client data live in your own workspace, never inside the plugin, so updating Foundry never touches them.
The rule that makes it safe near a real ad account
Agents are good at analysis and risky at spending. So the rule is enforced in code, not in a prompt. A live pause or budget change requires the confirm flag, fresh account data and a typed campaign key at an interactive terminal; budget changes also face a cumulative 20% cap over seven days. Runs without an interactive terminal are refused. The code notes that an agent with a pseudo-terminal could still answer the prompt, so keep permission prompts enabled for these commands. Campaign names and lead-form text are treated as untrusted data, because anyone with access to an account can write them.
Where it runs
- Codex: plugin and skill. A nightly CI job checks that Codex discovers it, and I ran a real session in which it loaded the skill and ran an audit.
- Claude Code: plugin with slash commands such as /foundry:audit and /foundry:report.
- OpenCode: Agent Skill, checked the same way in CI.
- ChatGPT: an offline Custom GPT pack for audits and calculators. I have not tested it inside ChatGPT's interface yet.
Built Arabic-first, for MENA
Triggers in Arabic, reports in Arabic, regional benchmarks, and the maths that matters there: cash on delivery, confirmation and delivery rates, VAT by country. Bundled playbooks, swipe files and worked examples cover COD operations, funnels, launches, and English and Arabic hooks.
The honest part
Foundry has a test suite and a CI matrix on Linux, Windows and macOS. A red-team pass found bugs in my own code, including a budget cap that the value "nan" slipped past; the repository includes a regression test for non-finite budgets.
What it has not done yet: run its live Google and Meta write paths against a real ad account. Those are covered by mocked tests. Start read-only.
Try it
Find Foundry in the registry: https://hol.org/registry/plugins/mohamedomar11111%2Fgrowth-architect-foundry Source (MIT): https://github.com/mohamedOmar11111/growth-architect-foundry
If you run an agency, run the audit on a test client and tell me what it got wrong. Launch stories like this one are how the directory gets useful.
