MCP and connectors for product marketers: connect one tool for one job

What a connector is

When you connect Google Drive, Slack or Notion in Claude, its available tools can read, and in some cases write, within the access granted to that connection. Anthropic's connector guide, rechecked September 16, 2026, describes connector support across Claude surfaces. In an ordinary unscheduled chat, Claude uses a connector when the conversation invokes it. A configured Cowork scheduled task may read connected sources between chats, so review its schedule and permissions separately.

Two other things get confused with a connector. Files you add to a Project are a static copy you replace by hand; the Project setup lesson covers that, and this article assumes you have a Project. A connector is access to the source system through the tools and scopes you granted. The other is a vendor that publishes an MCP for its own product. That is the vendor's dataset offered through the same protocol, and you evaluate it as you would any data source. Separately, Steve exposes its own data through a Steve MCP endpoint, which is a different direction of access.

What clicking Allow grants

A connector cannot exceed your permissions in the source service. If you cannot open the sales team's win-loss folder in Drive today, connecting Drive does not change that. The connector's tools, granted scopes and organization policy may expose less than your account can open, so inspect its actual tool and scope list.

Data direction is set per connector and per tool. The guide's own example is an admin who lets Claude search and summarise email while preventing it from sending any. Approval behavior depends on the surface, mode, tool policy and organization settings, so inspect the effective setting for each tool. Anthropic's remote MCP guide, rechecked September 16, 2026, lists custom connectors for Free, Pro, Max, Team and Enterprise and limits Free users to one. It states no numeric maximum for paid plans; directory connectors have separate availability. Team and Enterprise owners can also govern connector availability for their organizations.

A connector may expose permitted material through its enabled tools and scopes, so the sorting in the safe GTM data lesson applies to it as much as to a paste. If a folder holds material you would not paste into a chat, do not connect an account that can open it without first narrowing the connection.

Run this list before you click Allow:

• I have named the one job this connection is for and the one tool it needs.
• I have opened the connector's tool list and know which tools read and which write, post or send.
• I have checked the effective approval setting for each tool and disabled write access the job does not need.
• I know what this account can see, including shared folders and channels I do not normally open.
• I have confirmed the Project's visibility and connector policy, and I use private visibility when the material requires it.
• A date is in my calendar to reopen Settings and re-read what is still allowed.

The last item is our recommendation. Anthropic's documentation does not say connector settings preserve why access was granted. Record the purpose and review date yourself, and revisit any persistent approval you enabled.

Pick a job a connector improves

Good candidates share three features: the evidence lives in a tool you already open, the question is bounded, and you would read the original anyway. Three examples from product marketing work:

• Pull the competitor questions sellers asked in one Slack channel over the last two weeks, with a link to each thread.
• Find every internal document in Drive that states a claim about a named competitor, so you can check which claims still hold before they go into a claim ledger.
• Read the approved messaging page in Notion so a positioning draft starts from the approved wording.

Poor candidates: "monitor Klue" (a connector reads your tools, not a competitor's website), anything whose first step is Claude posting or editing, and any question one file drag would answer. Each connection is a standing permission you have to maintain, so give it a job you will repeat.

Connect, run once, open the source

1. Connect the one tool

   In Claude's connector settings, add the connector, sign in with the account whose permissions Claude should inherit, and open its tool list. Check: you can name each tool and say whether it writes.

2. Restrict before the first call

   Turn off the write tools in Search and tools, or decide to approve them per use. Check: the tools left on are enough for the job you named.

3. Ask the bounded question

   Give Claude the channel or folder, the date range, and a request for a link to every item it uses. Check: the answer cites items you can open.

4. Open the source anyway

   Open every cited item that supports a material claim and compare it with the answer. Check: the quoted text matches, the date is inside your range, and every material claim in the answer has a source you can reach.

5. Decide keep or disconnect

   If the answer held up and you will repeat the job, keep the connection and note the date. Otherwise disconnect it. Check: Settings lists only the connections you meant to keep.

What a broken connection looks like

When pasting is still enough

If the job is one document, drag it into the chat. If it is one Slack thread, paste it with the link and date. This manual path leaves no connector permission to maintain.

Choose a connector when the value of repeat access outweighs the work of maintaining its sources, permissions and approvals. Otherwise, paste the bounded material. When someone else will run the job, the workflow handoff lesson covers writing it down.

Where the manual version runs out, and what Steve does

Keep the manual setup while its operator can keep the sources, permissions and review dates current. If maintaining collection takes more attention than interpreting the evidence, evaluate a system built for collection.

A connector alone does not create a monitoring program. A configured scheduled task can query connected sources, but the schedule, reachable sources and review process still need to be set. Steve's competitive-intelligence product addresses competitor collection and battlecards. Steve's MCP page describes self-serve connection into Claude, ChatGPT, Cursor and other assistants, with tiers by competitors tracked. Check that page for what each plan includes.

Separately, hiresteve.ai/mcp is a different offer: public-source competitor data such as pricing pages, changelogs, reviews and news, exposed as an MCP for Claude, ChatGPT and other assistants, with plans, limits and what each includes stated on that page. Evaluate it with the checklist above, the same as any vendor MCP.