What each agent can and cannot do, and the rules that keep it from ever publishing without you.
Inside Marketing Swarm: six working agents (content writer, SEO strategist, social publisher, landing-page editor, paid-search copilot, reporter), two support agents (reviewer, image generator) and an optional builder agent, coordinated through a team chat and a review panel on one server. The agents produce; a separate publisher component with the only credentials acts on approved items, within caps enforced in code.
Drafts articles, pages and emails from a brief, your brand voice file, banned-language list and fact file. Flags any claim it cannot source rather than inventing one.
Refreshes keyword targets from live search-volume data, picks topics, writes briefs, audits internal links and metadata, and proposes technical fixes.
Turns each approved article into company-page posts, drafts variants, schedules on approval. Uses aggregate page metrics only; never member data.
Edits copy and images on paid landing pages on request. A code-level guard freezes form wiring, tracking tags, indexing flags and page structure.
Reads your ad account hourly, proposes negatives, pauses, bids and budget shifts with estimated impact. Applies nothing on its own.
Pulls analytics, search console, ad spend and the swarm’s own activity into a weekly summary with what moved and why. Read-only everywhere.
Checks every draft for banned language, reading level, link hygiene and unsourced claims before it reaches a human. Runs on a different model vendor than the writer.
Produces featured illustrations in your house style from the article, with text-free output enforced so nothing misspelled ships.
A coding agent scoped to one repository that ships approved structural changes with a second-vendor code review before every commit. Never touches live services directly.
One Linux server. Each agent runs as its own system user with access to only its own folder and keys. The review panel and the team chat are the only interfaces; there is no public API surface. Your AI provider keys are stored server-side, never in the browser, and model usage is billed to you directly.
Only the publisher component can post, schedule, edit a live page or change an ad. Agents produce proposals; the publisher acts on approvals. A compromised agent has nothing to send with.
Approve, edit, reject, snooze, from the team chat or the panel. Rejections record a reason the agents read on their next run.
Articles per day, posts per page per day, maximum ad budget, maximum CPC. The agents cannot see or change them.
Form wiring, tracking tags, indexing flags and page structure on landing pages are checked by a guard before a change can even be proposed.
The writer and the reviewer use different model providers; so do the builder and its code reviewer. One vendor’s blind spot doesn’t pass unchecked.
Web pages, search results and inbound messages are treated as content to summarise, never as commands. The agents’ instructions come from your files and your chat, nothing else.
AI provider keys run on prepaid balances with no auto-reload. If a run misbehaves, it runs out of money, not out of control.
| Item | What | Why |
|---|---|---|
| Server | One Linux VM, 4 GB RAM or more, any cloud | Everything runs here. We install; you own the login. |
| AI provider keys | Two or more vendors (prepaid balances) | Separate vendors for making and checking; spend hard-stops when the balance is empty. |
| Site access | Publishing access to your blog and landing pages | Used only by the publisher component, only on approved items. |
| Data access | Analytics, search console, ad account (read; ads also propose-only write) | Inputs for the SEO, paid-search and reporting agents. |
| Team chat | Any chat the panel can post into | Where the queue appears and approvals happen. |
| Your rules | Brand voice, banned words, facts, frozen page elements | Written with you in week 2; the agents’ only source of instructions. |
Security, permissions, what happens when a model misbehaves. Thirty minutes, on our own system.
Request a walkthrough