From landing pages and campaigns to analytics, operational memory and reporting. Each tool solves a specific problem; together they form a broader toolkit for day-to-day marketing work.
I did not set out to build five MCP servers. I built the first integration, found the next missing piece during real work, and repeated that process. Each tool still has one clear responsibility, but together they cover a complete review cycle.
The workflow can inspect campaign structure and performance to find wasted search terms, budget constraints, zero-conversion spend and tracking gaps.
A finding can be saved together with client context, earlier decisions and a concrete follow-up task instead of losing it in chat history.
The agent prepares a specific change and shows the exact scope. In the prototyping setup, nothing is sent to an ad platform until the responsible person confirms it separately.
The result can be compared against GA4 acquisition and conversion data, including cost and ROAS from linked Google Ads accounts.
Finally, metrics, charts and conclusions can become repeatable PDF or PNG deliverables rather than the same deck being rebuilt by hand.
I packaged the family for Claude Code, Codex and Cursor because I do not want the workflow tied to one model host. The servers run locally and the tools that can spend money or alter delivery use a separate confirmation step. Public repositories are available to inspect, but testing access is currently coordinated while credentials and integrations mature.
This was the first piece of the family. It can inspect accounts through GAQL, automate repeatable reviews and prepare changes for Search, Display and Performance Max from one agent session. Because it can alter budgets and delivery, every mutation has a two-phase confirmation flow.
GitHubThe family also includes a local workflow for the Meta Marketing API. Meta does not provide the comfortable sandbox I would want for campaign changes, so it shares the preview-first safety contract with the Google integration instead of treating confirmation as a sentence in the prompt.
GitHubCampaign-platform metrics are only half the story. I added a deliberately read-only GA4 integration for acquisition, conversion and channel performance. It can also retrieve linked Google Ads cost, CPC and ROAS through the GA4 Data API without granting another tool write access.
GitHubThis is the memory layer the other tools were missing. It stores client profiles, decisions, reviews, tasks and reusable marketing procedures as ordinary files. The files can be inspected, edited and versioned directly, carrying reasoning from one agent session into the next.
GitHubAnalysis is not finished when the agent prints a table in the terminal. I built Report Baby to turn structured data, HTML and reusable templates into PDF reports, PNG charts and metric cards. It gives the workflow an output that can actually be sent, reviewed and compared later.
GitHubHubSmith grew out of the other side of my marketing work: building and adjusting HubSpot landing pages. The browser editor is fine for content, but it is a poor place for understanding a whole theme, comparing modules or asking an LLM to make a careful change. HubSmith synchronizes themes and modules to a normal filesystem, renders a local preview and pushes the result back to HubSpot.
A coding agent can inspect the real module structure, follow imports, search every template and reuse conventions already present in the theme. Instead of describing a landing page through screenshots and pasted fragments, I can give the model the same source material I have.
The marketing MCP family is about helping an agent inspect, remember, measure and report on marketing work. PolyTrans is about multilingual publishing and editorial workflows. Both grew from practical work around the Trans.eu ecosystem, but they are not one pipeline and one does not depend on the other.
