Connectors and the marketplace: plugging Claude into your apps

The concrete flow: "pull the Q3 figure from my Drive doc"

Here is what happens end to end when you ask Claude to fetch something from Google Drive.

You type:

> "In my Drive, find the doc called 'Q3 Board Review' and tell me the net revenue retentionnet revenue retentionNet Revenue Retention measures the percentage of recurring revenue retained and grown from existing customers over a period, including upsell and expansion, net of downgrades and churn.Voir la définition complète → figure."

1. Claude recognizes it needs an external tool. Your prompt references your Drive, so Claude proposes calling the Google Drive connector. The first time, you go through an OAuth consent screen from Google, the same kind you have approved a hundred times. You grant Claude scoped access.

2. Claude searches, not browses. It issues a search to the Drive connector for "Q3 Board Review." The connector returns matching file metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). (titles, IDs), and Claude requests the contents of the one that matches.

3. Only that document's text enters the context window. Claude reads the returned content, locates the net revenue retentionnet revenue retentionNet Revenue Retention measures the percentage of recurring revenue retained and grown from existing customers over a period, including upsell and expansion, net of downgrades and churn.Voir la définition complète → number, and answers. Your other thousand Drive files were never opened.

4. You see the tool call. Claude.ai shows you that it used the connector and which document it pulled, so the action is auditable in the transcript.

That last point matters. Connectors are not a background daemon scraping your whole account. They fire on specific requests, fetch specific things, and show their work.

What data is shared, and what is not

This is the part people get wrong, so be precise.

What goes to Claude when a connector runs:

• The specific items Claude requests for your task. For the email example, that is the unread messages in the window you asked about (subject, sender, body). For Drive, the contents of the file(s) it opened.
• The metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). needed to find those items (search results, file names, message IDs).

What does NOT happen:

• Claude does not get a standing copy of your entire inbox or Drive. It pulls per request.
• Connecting Gmail does not mean every future chat silently reads your email. The connector is invoked only when a prompt warrants it, and you can watch it happen.
• Your connected data is not used to train Anthropic's models. Consumer and Team/Enterprise data handling on the Claude apps excludes your content from training by default. Verify the current terms for your plan in the Anthropic privacy and usage docs, because the exact defaults differ between consumer, Team, and Enterprise.

The honest caveat: once content enters the context windowcontext windowThe context window is the maximum amount of text (measured in tokens) a language model can process at once, including both the input prompt and the generated output.Voir la définition complète → for that conversation, it is part of that conversation. If you paste an unread email into a chat and then share that chat, the email goes with it. Connectors reduce manual copy-paste, but the data they pull still lives in the transcript. Treat a chat that touched real email or financial docs with the same care you would treat a forwarded thread.

Scoping access: grant the least you need

When you authorize a Google connector, the OAuth screen lists scopes. Pay attention to read-only versus read-write.

• A summarize-my-email task needs read access only.
• A "draft a reply and save it" task needs compose/write access.

Grant the narrower scope when the connector offers a choice. You can revoke any connector's access at two layers: inside Claude (Settings → Connectors → Disconnect) and at the provider (your Google Account's Security → Third-party access). Revoking at Google instantly cuts Claude off regardless of Claude-side state, which is the stronger control.

Connectors inside projects and skills

Connectors get more powerful when combined with the other Claude features you have already learned.

In a Project, you can enable specific connectors so every chat in that Project inherits them. A "Weekly Investor Update" Project with the Drive and Calendar connectors on means you can open a new chat and immediately say "pull this week's metrics doc and my Monday calendar, draft the update." The Project's custom instructions plus its connectors give Claude a consistent, tool-equipped workspace.

With Skills, connectors become the data source for a repeatable procedure. A Skill is a packaged set of instructions (and optionally code) Claude follows for a recurring task. A "Board Deck Prep" Skill might say: use the Drive connector to fetch the latest financial model, extract the three headline figures, format them as a summary table. The Skill defines the steps; the connector supplies the live data each time you run it.

This is the real unlock: connectors are not a standalone feature, they are the input layer for Projects, Skills, and Artifacts working together.

Custom connectors: when the marketplace is not enough

The certified marketplace covers common SaaS tools, but your company runs internal systems too. Because connectors are just remote MCP servers, you can add your own.

On Claude.ai (Team and Enterprise plans), an admin can add a custom connector by pointing Claude at a remote MCP server URL. Your internal tool, a customer databasecustomer databaseCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.Voir la définition complète →, an inventory APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète →, becomes available to Claude exactly like a marketplace connector, gated by your own auth.

A minimal remote MCP server that exposes one tool looks like this:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("inventory")

@mcp.tool()
def get_stock(sku: str) -> dict:
    """Return current stock level for a product SKU."""
    record = warehouse_db.lookup(sku)
    return {"sku": sku, "on_hand": record.quantity, "location": record.bin}

if __name__ == "__main__":
    mcp.run(transport="streamable-http")

Deploy that behind your authentication, register its URL as a custom connector, and now "how many units of SKU-4471 are in stock?" hits your warehouse in real time. The full spec and SDKs for building these live at modelcontextprotocol.io and in Anthropic's MCP docs.

Two things to keep straight:

• Marketplace connectors are certified and hosted; you just authorize them.
• Custom connectors are your responsibility for security, uptime, and scope. You own that server.

How this differs from the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → and Claude Code

You will hear "connector," "MCP," and "tool use" thrown around interchangeably. Sort them by surface:

• Connectors on Claude.ai are the no-code, UI-toggled version for end users. This lesson.
• The Messages API with tool use is where you, the developer, define tools and run the loop yourself in code. Different surface, same underlying idea of giving the model capabilities.
• Claude Code connects to MCP servers from your terminal so Claude can act on your local repo and dev tools.

The mental model that carries across all three: MCP is the protocol, and connectors are the friendly, certified, click-to-enable face of it inside the Claude apps. Learning to reason about scopes and the trust boundary here transfers directly to the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → and Claude Code surfaces you will see next in this path.

A realistic end-to-end task

Put it together. You want a Monday morning briefing.

1. Create a Project called "Monday Brief."

2. Enable the Gmail and Calendar connectors on it.

3. Add a custom instruction: "Every brief should list unread email needing a reply, then today's meetings with any prep needed."

4. Each Monday, open a chat and say: "Run my Monday brief."

Claude reads the unread mail (read scope), checks today's calendar, and produces the briefing as an Artifact you can edit. No manual copy-paste, the data pulled is visible in the transcript, and nothing left your accounts except what that task required.

Key Takeaways

• Connectors are certified, hosted MCP servers you enable from Settings on Claude.ai. They let Claude read live data from Gmail, Drive, Calendar, and third-party apps without any code from you.
• Data moves per request, not in bulk. Claude pulls only the specific items a task needs, shows the tool call in the transcript, and (by default on the Claude apps) your connected content is not used for training. Confirm the terms for your plan.
• Scope tightly and revoke at the source. Grant read-only when that is all the task needs, and remember that revoking at the provider (Google Account security settings) cuts access instantly and completely.
• Combine connectors with Projects and Skills to build repeatable, tool-equipped workflows like a Monday brief, rather than treating them as a one-off party trick.
• When the marketplace falls short, build a custom connector as a remote MCP server, but own its security and scope yourself. The same MCP model carries forward to the APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.Voir la définition complète → and Claude Code.