An assistant that finds wasted ad spend but cannot change the account has given you another task. In under two years, MCP helped turn that assistant from an adviser into an operator.
The sequence matters: a shared connection standard, support inside assistants, practical remote access, then tools that could write changes rather than just report problems. Each step removed a different piece of the copy-and-paste job that marketing kept calling optimization.
I care about the distinction because the expensive part was never getting another list of recommendations. It was getting the right changes into the account without handing an assistant permission to improvise with the budget. Here is the timeline of what changed, what it fixed, and where the remaining work went.
Before November 2024: the marketer was the connection
Before MCP, there was no common plug. The assistant lived in one tab, the ad platform in another, and the operator carried information between them. A normal review meant pulling a search terms report, cleaning it, pasting a chunk into Claude or ChatGPT, getting a list of suggested negatives, then adding them manually.
The assistant could give useful advice. Execution still belonged to the person with the browser tabs.
That gap mattered. Say you are spending $20k/month across three accounts: a review that ought to happen daily slips to weekly because moving the numbers takes too long. Better recommendations do not fix that bottleneck. They can just lengthen the queue.
I call that transfer work. It takes attention, but it is not the same thing as deciding where the money should go. The bottleneck was not ideas. It was hands.
November 25, 2024: an open specification makes connectors worth building
Anthropic announced MCP on November 25, 2024, with specification version 2024-11-05 and Python and TypeScript SDKs available from launch. Early partners included Block and Apollo on the data side, plus Zed, Replit, Codeium and Sourcegraph on the tooling side.
That list tells me who the first audience was: developers, not marketers. No reason to pretend an open specification immediately became a Google Ads operator.
The important change was the shared interface. Instead of building every connection around one assistant, developers could build an MCP server that compatible assistants could use. Same plug shape, less bespoke wiring.
For a PPC operator, that changes the useful question. Not “Does Claude know about ads?” but “Is there a connector that lets it work with my account?” Knowledge and access are different problems. MCP addressed the second.
A standard does not do the work. It makes the work worth building. That was the first turning point, not the finish line.
March 26, 2025: remote connections get a practical foundation
The 2025-03-26 revision added an OAuth 2.1 authorization framework and Streamable HTTP. Cloudflare documents Streamable HTTP as the standard remote transport introduced in March 2025.
In plain English, the connection no longer had to revolve around a server running on someone’s laptop. A server could live on the web, with a sign-in flow rather than a private key pasted into a conversation.
That matters when the account contains client spend and customer information. “Copy this credential into the chat” is not the onboarding experience I want to explain to a client.
Still, a sign-in screen is not a complete safety policy. It does not tell me which changes an assistant can make or who approves them. This milestone made remote access more practical; it did not settle governance.
Separate permission to connect from permission to change things. Marketing needed both, and they arrived as different pieces of the system.
May 1, 2025: Claude puts remote MCP inside the assistant
On May 1, 2025, Anthropic launched Claude Integrations, built on remote MCP servers. Launch integrations included Jira, Confluence, Zapier, Cloudflare, Intercom, Asana, Square, Sentry, PayPal, Linear and Plaid. Availability began in beta for Max, Team and Enterprise plans, then expanded to Pro by June 3.
Read that list as a marketer. Jira, Asana and Linear: work tools. Zapier: a bridge. Intercom, Square and PayPal: adjacent to the work. Not one ad platform.
That was the useful signal. Support inside a familiar assistant made the standard less abstract, but the assistant’s new hands reached for tickets and documents first.
For Google Ads practitioners, this proved a connection pattern, not a finished workflow. Assistant support and ad-account support were separate milestones. Before replacing any process, I would check the actual tools available, not the integration announcement.
June 18, 2025: identity and structured replies get stronger
The specification update dated 2025-06-18 made MCP servers OAuth resource servers with audience-bound tokens. It also added structured output and elicitation.
Those are unglamorous terms for useful changes:
- Audience-bound tokens tie credentials to their intended server.
- Structured output gives tools a defined shape to check rather than an answer buried in prose.
- Elicitation lets a server ask for missing information instead of proceeding without it.
For marketing, the direction matters more than the vocabulary. A system working around spend needs clear identity, checkable responses and a way to stop when information is missing.
None of that automatically makes a connector safe to operate an account. It gives its builder better pieces to work with. Look for those pieces in the implementation, not just “MCP-compatible” on the sales page.
November 25–December 9, 2025: consent improves and MCP gets a new home
The same timeline records a November 25 update adding icons and incremental consent, followed by MCP’s December 9 donation to the Agentic AI Foundation under the Linux Foundation. The foundation was co-founded with Block and OpenAI, with backing that included Google, Microsoft, AWS, Cloudflare and Bloomberg.
The announcement listed adoption across ChatGPT, Cursor, Gemini, Copilot and VS Code, alongside more than 10,000 public servers.
I read this as the point where MCP stopped looking like one assistant vendor’s integration project and started looking like shared infrastructure. Incremental consent also pushed the connection model toward a sensible idea: ask for permissions as they become necessary, rather than treating access as one enormous yes.
For practitioners, broader adoption made a connector less dependent on a single assistant interface. It did not make every server equally capable or trustworthy.
A shared standard reduces bespoke wiring. It does not remove the need to inspect what is connected. Plumbing still needs valves.
February–May 2026: official ad-platform servers arrive
The next useful change was not another assistant learning MCP. It was ad platforms supplying their own servers.
The official-server roundup records this sequence:
- February 2026: Amazon arrives.
- April 28: Google launches an open-source, read-only server.
- April 29: Meta launches a hosted open beta at
mcp.facebook.com/ads. - May 13: TikTok announces its server at TikTok World.
Those arrivals did not all mean the same thing. Meta shipped 29 tools across five categories, including reads and writes, with created entities paused by default. That last detail deserves more attention than the tool count. Creating an ad and putting it live are different decisions. Paused-by-default keeps a useful boundary between them.
Google’s official server documentation describes a self-hosted Python server with three read tools: listing accessible customers, running GAQL searches and retrieving resource metadata, plus four resources. Useful access, but not a reason to assume the assistant can carry out every recommendation it produces.
This is where “supports MCP” becomes too vague to help a buyer. One server reads. Another can write, but keeps new entities paused. Those are different operating models.
Check what the server can change, what stays paused and who says go. The logo on the connector does not answer those questions.
Mid-2026: the marketing shelf fills up, and setup becomes the next job
By 2026, the question was shifting from whether a server existed to which one to install first. A ranking based on 28-day search demand put Grafana at 529 impressions, Lemlist at 503, Google Search Console at 441 and LinkedIn Ads at 324.
I would not treat those figures as a verdict on connector quality. They show interest across dashboards, outreach, search and advertising. The demand was broader than ad-account operation alone.
That breadth creates its own problem. The same breakdown warns that 15 servers mean 15 separate authentication, setup and audit projects. A shelf full of connectors can put the operator back in the middle, joining data and checking separate logs.
I call that becoming the human join key. The tabs are newer. The transfer work is familiar.
If the architecture is still fuzzy, start with What Is an MCP Server? Your AI’s Hands, Your Rules, then the comparison in All-in-One Marketing MCP vs Separate Connectors: Stop Being the Join Key.
Start with one governed connection that sees spend and search together, not five clever ones that leave you stitching the work together. More connections are not automatically more automation.
July 28, 2026: simpler plumbing does not decide who approves
The protocol timeline records a move toward stateless operation and retirement of the old HTTP+SSE transport on July 28, 2026. That continues the infrastructure story: less dependence on the original connection model as MCP moves further into remote use.
It does not, by itself, answer the marketing question that matters most: who authorizes a change?
Read-only access leaves execution with the operator. Unrestricted writes hand too much discretion to the assistant. The useful operating pattern sits between them: enough access to finish the work, with explicit boundaries around what can happen.
My checklist is short:
- See what the assistant proposes to change.
- Know whether that action needs approval or falls within rules already set.
- Keep a log of what changed and why.
- Check whether the connector can undo the change.
If a connector cannot explain that workflow, I would not trust its write access. For the longer argument, read why read-only connectors leave you the work.
Start read-only, enable writes for one account, and keep creates paused until the log earns your trust. Transport is plumbing. Approval is an operating decision.
November 2026: the present is one governed connection
The useful destination is not an assistant with the longest tool list. It is an assistant that sees the relevant work, acts within rules and leaves a record I can inspect.
That is where groas sits. It connects ChatGPT, Claude, Cursor, Gemini and other AI agents to ad accounts and websites. It also collects information a plain account connector does not supply: competitor ads on Google and Meta and the funnels behind them, Reddit conversations, AI answers, and SEO data from Semrush and Ahrefs.
The operating controls matter as much as the coverage. Every change is logged with its reason and can be undone in one step. Autopilot stays off until I switch it on. groas Unlimited costs $19 a month, flat, for the whole team, with unlimited teammates, ad accounts and websites, every feature and all the data.
In Claude, I add it as a custom connector. In ChatGPT, I add it as a custom MCP server. The setup takes about a minute, without code, followed by sign-in rather than pasting credentials into a prompt.
The reason to start here is the lesson running through this timeline. Separate connectors can leave the operator responsible for joining spend, search and execution. One governed endpoint covering Google Ads, Bing, Meta, ChatGPT Ads, SEO and AI search visibility removes that stitching job.
The best marketing MCP is not the one with the longest tool list. It is the one I can trust with write access. That is a higher bar than producing a convincing answer.
Where I think the line points next
Three things matter more to me than another connector announcement:
- Approvals become the product. Paused-by-default is a useful start. The stronger proposition is setting rules once, then knowing which actions can proceed and which must wait.
- Cross-platform context stays valuable. Official servers serve their own platforms. Someone still needs to put search spend next to search visibility without making the marketer carry the data between them.
- Logs become the invoice. A recommendation sounds useful. A record of what changed, why it changed and how to undo it lets me inspect whether the machine earned its keep.
That is the line from the open specification to the present: connection first, execution next, accountability throughout. The point was never to replace copy and paste with a more impressive way to ask for advice.
Copy and paste was the job. Now the job is setting the rules and checking the math. Give the assistant hands, keep the approval, read the log.

