Kadabra's Shift to Claude as the Front Door

In 2026 Kadabra moved much of its daily work out of dashboards and into one conversation. Forty-five connectors, 12 of them built in house, link Claude to our CMS, our data and operations systems, SEO tools, project boards, email and reporting. Our people ask for data, analyze it and act on it in a single place. Here is what changed, what broke, what it took to do safely, and why it fits how Kadabra has always worked.

The old way: five tools and a spreadsheet

A routine question used to look like this: why did one of our guides lose organic traffic last week?

Answering it meant Search Console for clicks, a rank tracker for positions, our data systems for registrations and deposits, the CMS to see what changed on the page, and a spreadsheet to stitch it all together. Then a ticket. Then a message to the person who could actually fix it.

Every one of those tools is good at its job. The cost sat in the gaps between them: logins, exports, copy and paste, and the hour that passed before anyone could start thinking about the answer. Across hundreds of sites in 30+ markets, that hour repeats all day, every day. It is also the part of the job nobody joined Kadabra to do.

What changed: the interface moved, the tools stayed

We did not replace our tools. We stopped opening them.

Every system we rely on is now reachable from Claude through a connector: a small server that speaks the Model Context Protocol (MCP), the open standard that lets an AI assistant read from and act inside other software. We run 45 connectors. Most come off the shelf from the Claude connectors directory and cover project management, email, calendars, documentation and search data. Twelve we built in house, because no vendor was going to build them for an iGaming affiliate. Those reach our CMS estate, our data and operations systems, SEO auditing, image generation and reporting.

The traffic question is now one prompt. Claude pulls the clicks, the positions, the conversions and the page's edit history, lines them up and proposes a cause. If the fix is a title, a section or an internal link, it drafts the change in the CMS and opens the ticket. A person approves it.

	Before	Now
Ask	Log into five tools, export CSVs	One prompt pulls from search data, rank tracking, our data systems and the CMS
Analyze	Pivot tables, usually the next day	Claude joins the data and explains what moved, in the same conversation
Act	Write a ticket, hand off, wait	Draft the fix, update the board, publish the report, with a person approving

Claude became the place where we ask, analyze and act. The tools became the plumbing behind it.

What it looks like on a normal day

• SEO. A specialist audits a page, checks its archived history to see what changed and when, and gets internal link suggestions with anchor text, without leaving the chat.
• Content. Writers start from briefs built on live search data instead of a blank template.
• Data. Product and BI ask our data systems questions in plain language instead of waiting for a custom report, then publish the answer as a shareable dashboard.
• Performance. An automated daily review of affiliate performance tells the team what moved and why, before the first coffee.
• Operations. Agents running on the same connectors handle repetitive pipelines end to end, such as link prospecting, vetting and monitoring, and hand people the decisions.

Usage spread faster than we planned for. Dozens of people across the company now work through our connectors, on well over a hundred sites, making tens of thousands of tool calls. The change we care about more is harder to count: SEO, Growth and HR now build their own tools instead of queueing requests for developers.

What broke: demand outran the build

We expected the hard part to be adoption. It was the opposite.

Once people could reach their systems from one window, they used the connectors to the max. Real work exposed gaps no spec had predicted: a missing filter, a report format a team needed that week, an action that only made sense once people were living in the read tools. More than once we had to add features fast because a team was already waiting on them.

The lesson: ship connectors small and let real usage decide what comes next. The most useful roadmap we have is the list of things our people asked Claude to do that it could not do yet.

Why now, and why it is not a bet on one vendor

Two years ago this setup would have been fragile and proprietary. Three things changed.

The plumbing became a standard. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a new fund under the Linux Foundation co-founded with Block and OpenAI and backed by Google, Microsoft and AWS. By then there were more than 10,000 active public MCP servers, and ChatGPT, Gemini, Copilot and Cursor all supported the protocol. The connectors we build are not locked to Claude.

The know-how became portable. On October 22, 2026 Anthropic donated Agent Skills to the same foundation. More than 70 agent platforms support the format, and nearly 280 million skills have been installed. Our skills library, the playbooks that teach Claude how Kadabra does SEO, is written in it.

The industry is converging, from the opposite direction. Gartner predicts that 40% of enterprise apps will include task-specific AI agents by the end of 2026, up from under 5% in 2025. That is an agent inside every app. We went the other way: one assistant in front of every app. For a company with dozens of tools and hundreds of sites, one front door that knows the whole building beats forty smart rooms that only know themselves.

We are honest about the dependency. Claude is our front door today, and we chose it because it works best for us. Because connectors and skills are open standards, the integration work we have done would move with us if that ever changed.

What we had to get right

Giving an assistant write access to live sites is a different risk from giving it read access to a dashboard. Four rules came out of that.

1. Least privilege, per person. Each user reaches only the sites and tools their role needs. External partners get scoped access, never a shared login.
2. A person approves anything public or destructive. Publishing, deleting, DNS changes and outbound email need explicit confirmation.
3. Everything is logged. Every tool call through our gateway records who did what, where and when. That log is also how we see what people actually use and what they still need.
4. Connectors are reviewed like vendors. We keep an inventory of every connector, its scopes and its owner, and remove what nobody uses. That review is part of a company-wide security audit running now.

None of this slowed adoption down. It is what let management say yes. In our experience, permission decides how fast a company moves with AI far more than budget or technology does.

Why it fits Kadabra

Our vision has always been to help users make better decisions online. This shift gave our own people the same thing: the data and the means to act on it, in one place.

It also maps onto the values we wrote down long before MCP existed:

• Dynamic. We say we are not afraid of change because our market never stops changing. Moving the whole company's way of working inside a year is that value tested for real.
• Result-driven. Our loop is plan, execute, measure, repeat. It used to stall at "measure", because measuring meant exports. Now measuring is a question.
• Team players. A skill is a shared playbook. When one SEO specialist writes down how they vet a link, every colleague's Claude does it the same way the next morning.
• Trust. Each person owns their project. Connectors let them act on it directly instead of filing a request and waiting.

We spread it the way we spread everything else: people teaching people. In our AI Academy, colleagues present real case studies from their own work to the whole company. At our 2026 hackathon, teams of marketers, SEO specialists, analysts and developers built working AI tools in a single day. It is still all about the people. They just spend less of their day moving data around.

What comes next

The next step is fewer prompts, not more. The connectors that answer a question on demand are starting to run on a schedule: daily performance reviews, link checks and monitoring jobs that flag problems before anyone has to go looking. People set the goal and review the result.

If you run a marketing or affiliate team, one test is worth running this week. Pick a routine question and count how many tools it touches before anyone can start thinking about the answer. That number is the size of your opportunity.

If this is the kind of place you want to build in, we are hiring.