The first thing most teams do after buying Claude seats is open the Connectors screen. It is a wall of logos: Gmail, Google Drive, Slack, HubSpot, Notion, Asana, Microsoft 365, and dozens more from the directory. The natural instinct is to connect everything and see what happens. This guide is the case for doing the opposite. Claude connectors are less a technical decision than a permissions decision, and the teams that get real value from them are the ones who turn on a few, on purpose, with everyone understanding what just changed.
What Claude connectors actually are
Anthropic's own definition is the clearest one: connectors “let Claude access your apps and services, retrieve your data, and take actions within connected services.” Under the hood they run on the Model Context Protocol (MCP), an open standard Anthropic created so that AI tools and business apps can talk to each other in one consistent way.
Two things in that definition deserve a second look. Retrieve your data means Claude can read: search your Drive, pull a thread from Gmail, look up a deal in the CRM. Take actions means Claude can also write: draft and send an email, create a task, update a record. Reading is where most teams should start. Writing is where the rollout decisions live.
Connectors come in two flavors:
- Prebuilt integrations that Anthropic maintains for the big platforms: Google Drive, Gmail, Google Calendar, GitHub, Slack, and Microsoft 365. These work as soon as you sign in.
- Directory connectors built by third parties and listed in the Connectors Directory, which carries both “verified” and “community” entries. Anthropic is careful to say that verification is not a security audit, and that distinction matters more for a business than for an individual user.
There is a third option, custom connectors that point at any remote MCP server, and we will come back to why most small teams should hold off on those.
What data can Claude access through connectors?
This is the question owners actually want answered, and the answer is reassuringly boring. Claude inherits each person's permissions from the connected service. In Anthropic's words: “If someone can't access a specific file, channel, or record in the source system, the connector can't reach it from Claude either.”
So a connector does not open new doors. It walks through the doors that are already open for that person. If your Drive sharing is tidy, Claude's access is tidy. If your Drive sharing is a decade of “anyone with the link,” a connector will faithfully reflect that, and the fix is in Drive, not in Claude.
A few other facts worth having on hand when someone on the team asks:
- Data in transit is encrypted, per Anthropic's support documentation.
- Third-party services on the other end of a connector process data under their own terms, which Anthropic notes may be located outside the United States. If a client contract has data-residency language, check it before connecting that client's systems.
- On Team and Enterprise plans, an Owner enables a connector for the organization, and then each person signs in with their own account. Nobody inherits a colleague's login.
The connectors most small teams should start with
Every business is different, but the first wave looks similar almost everywhere. Pick the ones that match where your team already spends its day, and skip the rest for now.
- Your email and calendar (Gmail and Google Calendar, or Microsoft 365). This is the highest-value, lowest-risk starting point. First thing to try: “Summarize every thread from this client in the last two weeks and list anything I owe them.”
- Your file store (Google Drive or SharePoint/OneDrive). Claude searching your own documents beats Claude guessing. First thing to try: ask it to find the last three proposals you sent in a given service line and tell you how the pricing differed.
- Slack, if that is where decisions happen. First thing to try: “What did we decide about the onboarding checklist in #ops last week?”
- Your CRM (HubSpot is the most common in the shops we work with). Start read-only in practice: ask about pipeline, stale deals, and last-touch dates before you let it edit anything.
- Your project tool (Asana, Linear, Notion, or similar). Useful for status roll-ups and for turning meeting notes into tasks, which is the first write action most teams are comfortable with.
For agencies there is one more consideration: the tools your clients live in. Connecting a client's HubSpot or Drive is a conversation to have with the client first, in writing, even when the contract technically allows it. That is not a legal note so much as a relationship one.
What to leave off, for now
Not forever. Just until the first wave is settled and the team trusts what it sees.
- Custom connectors to servers you did not build or vet. Anthropic's guidance is blunt: “Only connect Claude to servers built and hosted by organizations and applications you trust,” because “malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions.” On Team plans only Owners can add custom connectors, which is a sensible default. Keep it that way.
- Anything that touches money or billing with write access. Accounting, payments, and payroll connectors are fine to read from. Turning on write actions is a decision for month two, with approval steps in place.
- Community connectors with no identifiable vendor behind them. The directory label tells you it exists, not that anyone is accountable for it.
- Systems whose permissions are already a mess. Because Claude mirrors source-system access, a connector will surface that mess quickly. Clean the sharing first; then connect.
Rolling connectors out on a Team plan without a mess
This is the part that gets skipped, and it is the part that determines whether people trust the tool six months from now. A rollout that works for a 10-to-50-person team tends to look like this:
- The Owner enables the first wave, not the whole directory. Team plan members who cannot enable connectors themselves see a Request button in the directory, and requests land in the Owner's organization settings. That queue is a feature: it tells you what people actually want, one connector at a time.
- Set write actions to “needs approval” at the start. Owners on Team and Enterprise plans can limit which actions a connected service can take, choosing among always allow, needs approval, or blocked for each tool. Anthropic notes these restrictions only narrow what the source system already permits; they never expand it. Start narrow.
- Pair every connector with one named workflow. “We connected Gmail so account managers can prep for client calls” gives people a reason to try it and a way to judge whether it helped. “We connected Gmail” gives them nothing.
- Tell people what changed, in one message. Which connectors are on, what Claude can and cannot see, and who to ask. Most of the anxiety about AI reading company data comes from not knowing the answer to that middle question, and the answer is usually calming.
- Review at 30 days. Which connectors got used, which requests came in, which write actions people kept approving without reading. That last one is your signal to move a tool from needs approval to always allow.
A note on Cowork
If your team is using Claude Cowork for scheduled or longer-running work, the same connectors apply; Cowork has full MCP support. The difference is that Cowork often works while nobody is watching, which makes the approval settings above more important, not less. A connector you would happily let Claude use in a live chat deserves a second look before it runs on a schedule at 6 a.m.
Where to start this week
Pick two connectors from the first-wave list, one that reads email or files and one that reads your CRM or project tool. Enable them for the organization, set write actions to need approval, and give each one a single named job. Then leave everything else off until the team asks for it by name.
If you would rather have someone sit with your team through that first month, from the permissions review to the 30-day check-in, that is the work Cedar & Loam does. Proudly based in Kansas City, working with agencies and small firms wherever they are. Say hello.
Sources
- Use connectors to extend Claude's capabilities, Anthropic Help Center
- Connectors overview, Claude documentation
- Connectors directory, Claude documentation
- Get started with custom connectors using remote MCP, Anthropic Help Center