What Anthropic actually shipped this week: plugins, connectors, and an MCP app for finance

Anthropic put out a bundle of releases on 5 May 2026: ten new plugins for Cowork and Claude Code, integrations across the Microsoft 365 suite, a set of new data connectors, and an MCP application aimed at financial services and insurance firms. The day before, on 6 May, they raised usage limits and disclosed a compute deal with SpaceX.¹

That's a lot of nouns in one paragraph. If you're trying to figure out what's actually new, versus what's a rename of something that already existed, this is for you. I'm going to walk through the shape of the announcement, because the four pieces fit together in a way that isn't obvious from the headlines.

The pattern: Anthropic is building a distribution layer

The problem this solves. A model on its own is useless to a knowledge worker. To do anything real, it needs to read your email, see your calendar, query your data warehouse, open a ticket, post to Slack, and remember what it did last Tuesday. Every one of those is a separate integration. For a year, the question facing every AI vendor has been: who builds those integrations, and how?

There are basically three answers. The vendor builds them all in-house (slow, doesn't scale). The customer builds them (expensive, only the biggest customers can). Or there's a standard protocol that lets anyone build them once and everyone use them. Anthropic has been betting on the third option since late 2024, when it published the Model Context Protocol, MCP, as an open spec.

MCP, in one sentence. It's a standard way for an AI assistant to talk to a tool or data source, so that any MCP-compatible assistant can use any MCP-compatible tool without bespoke wiring.

Think of it as USB for AI integrations. Before USB, every peripheral had its own connector and its own driver story. After USB, you plug it in and it works. MCP is trying to be that, for the connection between language models and the outside world. The metaphor breaks down, MCP carries structured tool calls and resource references, not bytes, but the shape is right: one protocol, many endpoints.

What Anthropic shipped this week is the first version of that vision where the pieces are visibly connecting into a product layer rather than sitting around as a spec.

The four pieces, and how they fit

Plugins (ten of them, for Cowork and Claude Code). A plugin, in this context, is a packaged capability that extends what Claude can do inside a specific surface. Cowork is Anthropic's workspace product; Claude Code is their developer-facing coding assistant. A plugin might be "run this linter", "open a pull request", "summarise this meeting transcript". The mechanism underneath is almost certainly MCP, plugins are how MCP servers get surfaced to end users without those users having to know what an MCP server is.

This matters because it's the first time Anthropic has shipped a plugin catalogue rather than a protocol. Protocols are for developers; catalogues are for everyone else.

Microsoft 365 integrations. Outlook, Teams, Word, Excel, SharePoint, OneDrive. The full suite. For Cowork users, this means Claude can now read your email, see your calendar, edit your documents, and pull from your shared drives, without anyone writing custom code.

This is the integration that turns Cowork from "a chat window with a smart model" into "an assistant that can act on the systems you actually use at work". Most enterprise knowledge work happens inside Microsoft's surfaces. Until you can read and write to those surfaces, you are a toy.

Data connectors. These are the pipes to structured data sources, warehouses, CRMs, ticketing systems. The connector reads the schema, exposes the right tables and fields to the model as MCP resources, handles authentication and permissions. Each connector is a small, focused piece of software whose only job is to make one data source legible to an AI assistant.

The interesting design choice here is that connectors are not the same as plugins. A plugin is an action (do this thing); a connector is a source (read from here). Splitting them keeps the surface clean: the model knows what it can do, separately from what it can see.

The MCP application for financial services and insurance. This is the piece that's harder to read from a single announcement, because "MCP application" is a new product category. My read: it's a vertical-specific bundle, a set of connectors, plugins, and probably some pre-built workflows, packaged for one industry. Financial services and insurance have heavy compliance requirements, specific data sources (Bloomberg, claims systems, regulatory filings), and a need for auditable AI behaviour. Shipping a bundle aimed at them is Anthropic acknowledging that horizontal tools don't quite fit vertical work.

A worked example

Imagine you're an underwriter at a mid-sized insurer. A broker emails you a new submission. With this stack, here's what becomes possible:

Claude reads the Outlook email (Microsoft 365 integration), extracts the risk details, queries your policy admin system for the client's history (data connector), pulls comparable historical losses from your data warehouse (data connector), runs them through your rating model (plugin), drafts a response in Word with the quote and the rationale (Microsoft 365 integration), and logs the whole thing to your audit system (plugin).

None of those steps individually is new. What's new is that they're composable through one protocol, addressable from one assistant, and packaged together as something an IT team can deploy rather than something a research team has to build.

Trade-offs

You're trusting one protocol with a lot. If MCP has a security flaw, every connector inherits it. Anthropic has been careful here, MCP has explicit permission scopes, human-in-the-loop confirmations for sensitive actions, and audit logging, but the surface area is now large enough that this is a meaningful concern, not a theoretical one.

Vendor concentration. Anthropic is building both the protocol and the application layer on top of it. MCP is an open spec, and other clients (including some from OpenAI and Google) are starting to support it, but right now the highest-quality MCP experience is inside Anthropic's products. Open protocols can still produce concentrated markets, see also: email.

Plugins are only as good as their underlying systems. A plugin that calls Salesforce can't be faster or more reliable than the Salesforce API. When something breaks, the debugging story crosses three vendors.

What to watch

The 6 May disclosure of higher usage limits and the SpaceX compute deal is the other half of this story. Plugins and connectors generate a lot of tokens, every tool call round-trips through the model, and a single underwriting workflow like the one above might burn tens of thousands of tokens. If Anthropic is shipping the integration layer, they need the compute to back it. The two announcements, one day apart, are the same announcement.

I'd also watch whether the "MCP application" framing extends to other verticals. If healthcare, legal, and logistics versions show up over the next two quarters, this is a product strategy. If it stays at financial services, it was a single deep customer engagement that got generalised.

And the open question I haven't seen anyone answer clearly: when does Microsoft itself ship a competing plugin layer for Copilot, using its own protocol rather than MCP? That's the move that decides whether MCP becomes the standard or becomes one of two standards.

Footnotes and links

Further reading

• [MCP spec]: modelcontextprotocol.io, the protocol itself, with a short and readable overview
• [Why MCP matters]: Simon Willison's running notes on MCP adoption are the best ongoing commentary

¶