Claude Code stopped being just a coding tool this year. Anthropic shipped a plugin system that lets it connect to hundreds of external business tools through the Model Context Protocol, and we think most operators are reading the wrong headline about it.
What actually shipped
On October 9, 2025, Anthropic moved Claude Code plugins into public beta. A plugin bundles slash commands, subagents, MCP servers and hooks into one installable package, and it connects Claude Code to hundreds of external tools and data sources through the Model Context Protocol, an open standard described at modelcontextprotocol.io. You install one from a marketplace with a single command, point it at a system you already pay for, and Claude Code reads and acts on that system directly instead of you copying data between browser tabs.
That is the official framing, and it is accurate as far as it goes. But we run a 40-person operation and build these systems for other operators every week, and the part that matters to us is not the installation mechanic. It is what happens once the wiring is done and the agent is sitting on top of your CRM, your ad account, your data provider and your inbox at the same time.
The workflow we kept coming back to
The clearest example in front of us this week connects Claude Code to Clay, the B2B data platform, through its own plugin. Once authenticated, you can ask Claude Code in plain language to source a list of leads matching a description. It enriches them with company and contact data, writes personalized outreach copy, and hands the whole thing back as a table ready to load into a campaign tool. No new dashboard to learn. You describe the outcome and the agent works the API calls.
The mechanism underneath is what we would point any operator toward. Clay enriches records through a waterfall. It checks its top data provider first, and if that provider does not return a match, it falls through to the next one, and the next, until it finds a hit or runs out of sources. Clay's own claim, stated directly on its waterfall enrichment page, is that this routinely triples customer data coverage compared to relying on a single provider. That is not our number. It is theirs, and we are citing it because the mechanism, not the vendor, is the useful idea.
Here is a simple way to see the shift:
| Old way | Claude Code plugin way |
|---|---|
| Log into 4 to 6 separate tools per task | One conversation, agent calls the tools |
| Copy and paste data between tabs | Data moves through MCP directly |
| Single data provider, lower match rate | Waterfall across providers, higher match rate |
| Manual, ad hoc outreach copy | Copy generated from your actual context files |
Why this is bigger than lead generation
Lead sourcing is the demo everyone shows because it is easy to film. But the same pattern applies to anything with an API: a support inbox, an ad account, a project tracker, a calendar, a bookkeeping system. Once a tool is reachable through MCP, an operator who never wanted to learn that tool's interface can still direct work inside it, in the same conversation where they are directing work in five other tools. We build automation systems for clients precisely because the bottleneck was never the individual tool. It was always the switching cost between tools, and the fact that only one or two people on a 15-person team ever learned all of them well enough to move fast.
The question nobody in the demos is asking
Every one of those actions costs something. Clay enrichment runs on credits, ad platforms bill on spend, and a plugin that can send email or place an order needs a limit before it needs a feature. If you connect five tools to an agent and give it no boundaries, you have not built an operator. You have built an intern with a company card and no manager. We treat every plugin connection the way we would treat onboarding a new hire: scoped access, a spending ceiling, and a log of what it did and why. That governance layer is exactly what we scope in an AI Concierge assessment before any tool gets wired to anything.
The bottleneck is not the wiring, it is the documentation
The demo that convinced us of this had one detail most people skip past. Before any lead got enriched, the operator had already loaded Claude Code with a business profile, case studies, an FAQ, proof points and website copy. Without that context, the agent could source data all day and still write outreach nobody would answer. The output was only as sharp as the documented business sitting behind it.
That is the exact gap we see in almost every 10 to 50 person company we assess. The knowledge that would make an agent genuinely useful, pricing logic, objection handling, case studies, standard operating procedures, lives in someone's head or scattered across old decks and Slack threads, not in a form any tool can read. This is precisely the problem a Second Brain is built to close: one structured, owned repository of how your business actually runs. When you connect a plugin, the agent has something real to work from instead of a blank slate. Clients keep 100 percent of that code and those files. It is not a black box you rent.
What we are building into client systems this quarter
Three moves, in order, for any operator looking at this shift:
- Audit before you connect anything. List every tool your team logs into weekly and rank them by how much manual copying happens between them. Start your first MCP connection with the worst offender, not the flashiest one. Our revenue leak heatmap is a fast way to see where the switching cost is actually bleeding hours.
- Write the context down before you wire the tool. Business profile, offer, FAQs, case studies, pricing rules. If a new hire could not do the job from those documents alone, neither can an agent connected through a plugin.
- Put a spending and access limit on every connection before it goes live. Treat it like payroll, not like a browser extension.
We have seen the same pattern across the client builds we run: the first working system goes live in days to weeks, not months, once the context is documented and the access boundaries are set. The agent runs the checks 24 hours a day once it is live, which a person on a 40-hour week structurally cannot match.
The numbers that actually hold up
To be precise about what is verified and what is directional: Claude Code's plugin system reaches, in Anthropic's own words, hundreds of external tools and data sources, and it left beta status as a public feature on October 9, 2025. Clay states its own waterfall enrichment routinely triples data coverage against a single provider, which is a vendor claim we are citing, not one we measured ourselves. On our side of the table, the approved facts we work from with clients are steadier. A five-minute response window is the practical threshold before an inbound lead goes cold. Most 10 to 50 person teams are leaking 10 or more hours a week to admin that a connected agent should be doing instead, and an AI Concierge assessment starts at 999 euros, credited in full to the build if you move forward.
None of that requires believing every demo. It requires deciding which tool in your stack is costing you the most switching time this month, documenting what a competent employee would need to know to run it well, and only then asking an agent to pick up the work. We write more of this as we test it against real client systems on the luup blog. If you want the audit done on your own stack instead of a generic checklist, that is what the assessment is for.