# Claude for business: plans, connectors and connecting your data with MCP

Claude for business means Anthropic's Team and Enterprise plans plus the Claude API. The plans give employees a shared workspace with admin controls, and connectors let Claude read and act in company tools with each user's own permissions. Systems without a ready-made connector are reached through a custom connector, which is a remote MCP server that your company builds or commissions.

By Shai Goldenberg. Updated: 2026-11-14.

## What Claude is for a company

Claude is Anthropic's AI model and the set of products built on it. A company meets it in three forms.

- **The Claude apps.** A chat workspace on web, desktop and mobile, where employees write, analyze documents, search connected tools and run longer tasks. This is what the Team and Enterprise plans cover.
- **Claude Code.** A coding agent that works in a developer's terminal and repository. Anthropic's pricing page lists it as included in the Team and Enterprise plans.
- **The Claude API.** Access to the models for building your own products, such as a support chatbot or an internal agent. Billed by usage.

The first two are tools your people use. The third is a component your engineers build with. Most companies end up using more than one, and the decisions are different for each.

## Plans for business use

The facts below are from Anthropic's pricing page as of November 2026. Plans and prices change, so check the page before you budget.

| | Team | Enterprise |
|---|---|---|
| Size | 2 to 150 seats | 20-seat minimum |
| Price | Per seat, billed monthly or annually. See the current figure on Anthropic's pricing page | Per seat plus usage at API rates, agreed with Anthropic sales |
| Connectors | Yes | Yes |
| Claude Code | Yes | Yes |
| Admin controls for connectors | Yes | Yes |
| Single sign-on | Yes | Yes |
| SCIM provisioning | Not listed | Yes |
| Audit logs | Not listed | Yes |
| Custom data retention controls | Not listed | Yes |
| Compliance API | Not listed | Yes |

How to read the table: Team covers a company that wants shared billing, SSO and connectors. Enterprise adds the controls a security or compliance team asks for: automated user provisioning and removal, a record of activity, retention rules, and programmatic access for compliance tooling. If your security review requires audit logs, that decides the plan.

Individual plans exist too. They are personal subscriptions without central administration, so they are a poor base for company data.

## Connectors: how Claude reaches your tools

A model knows nothing about your company until something connects it. In Claude, that something is a connector.

Anthropic's documentation describes connectors as letting Claude retrieve your data and take actions within connected services. Its examples include searching files in Google Drive, sending messages in Slack and creating issues in Linear. Available connectors are listed in a directory, and each entry says whether the connector reads, writes or both.

Four facts from the documentation matter for a rollout:

1. **An owner enables, each member connects.** On Team and Enterprise plans, an Owner or Primary Owner enables a connector for the organization. Enabling does not grant anyone access. Each member then signs in to the service with their own account.
2. **Permissions are inherited.** Anthropic states that Claude inherits each person's permissions from the connected service. A file a user cannot open stays closed to Claude when that user asks.
3. **Owners can narrow what a connector does.** Actions can be set org-wide to always allowed, needs approval or blocked. A common setup is to allow search and reading and block sending or editing. Restrictions only narrow access. They never grant more than the source system allows.
4. **Requests are controlled.** Whether members can request new connectors is a setting, and its default differs between Team and Enterprise.

This model is sound, and it places one requirement on you: permissions in the source systems must be correct. If half the company can open the finance folder today, Claude will make that visible fast. Many rollouts begin with a permissions cleanup.

## Custom connectors and MCP servers

The directory covers popular software. It does not cover your ERP, your internal admin system or your data warehouse. For those, Claude supports custom connectors, and a custom connector is a remote MCP server.

MCP, the Model Context Protocol, is an open standard for connecting AI applications to tools and data. An MCP server wraps a system and describes its capabilities in a form the model can discover and call. The full explanation is in [what MCP is](https://goldenroz.com/blog/what-is-mcp/), and the build steps are in [how to build an MCP server](https://goldenroz.com/blog/how-to-build-an-mcp-server/).

What Anthropic's documentation says about custom connectors, as of November 2026:

- **Who adds them.** On Team plans, Owners and Primary Owners. On Enterprise plans, also members with a custom role that includes library management.
- **Where the connection comes from.** Claude connects to your remote MCP server from Anthropic's cloud infrastructure, not from the user's device. A server on a private network or behind a firewall will not connect unless you allowlist Anthropic's IP addresses.
- **Authentication.** OAuth is the typical method, so each member signs in as themselves. A server can also use a fixed credential sent as a request header, which Anthropic says is stored encrypted.
- **The shared credential trap.** If a connector uses one shared credential, everyone who uses it gets the same access. Anthropic's guidance is to use a credential limited to what members should reach.
- **Trust.** Custom connectors are not verified by Anthropic. Its documentation warns to connect only to servers you trust, and notes that a malicious server can hide instructions in tool inputs or outputs.

### What a good internal MCP server looks like

The protocol standardizes the connection. What the connection may do is decided by whoever builds the server. A server fit for company data has these properties:

- **Read-only by default.** Write tools are added one at a time, with limits.
- **Per-user permissions.** The server acts as the person asking, through OAuth, never as a shared admin account.
- **No raw queries.** The model picks from approved queries or narrow tools. It does not write SQL against production.
- **PII masking.** Personal data the task does not need is removed before it reaches the model.
- **An audit log.** Every tool call is recorded with the user, the arguments and the result.
- **Injection defenses.** Text returned from systems is treated as data, never as instructions.

These are the defaults in Shai's on-prem MCP gateway, which lets Claude or ChatGPT answer questions from ERP, CRM, SQL and Microsoft 365 systems. It has 13 connector types and installs at the customer with Docker Compose. Our [MCP server development](https://goldenroz.com/mcp-server-development/) service builds custom servers on the same principles.

## Data handling questions to ask

Take these to your security and legal teams, and answer them from Anthropic's current documentation and your contract, not from a blog post.

| Question | Where the answer is |
|---|---|
| Is our data used to train models? | Anthropic's privacy documentation states that by default it does not train on inputs or outputs from commercial products. Feedback that users submit is the documented exception, and owners can disable feedback |
| How long are conversations retained, and can we set it? | Plan features and your agreement. Custom retention controls are listed under Enterprise |
| Who inside our company can see what? | Connector permissions and your source system permissions |
| What can connectors change, not only read? | The org-wide action settings for each connector |
| Is there an activity record for audits? | Audit logs and the Compliance API are listed under Enterprise |
| Where does data from our internal systems go? | The design of your MCP server: what it returns, what it masks, what it logs |
| What happens when an employee leaves? | SSO on both plans. Automated deprovisioning through SCIM is listed under Enterprise |

The last-but-one row is the one you control fully. Anything an MCP server returns enters the model's context. The decision about what leaves your network is made in the server's code.

## A rollout plan

**1. Pick the first use cases.** Two or three, each with an owner and a way to tell whether it worked. "Sales answers account questions without asking finance" is a use case. "Use AI" is not.

**2. Fix permissions in the source systems.** Before enabling connectors, review who can see what in the drives and tools you will connect.

**3. Choose the plan from the security requirements.** Audit logs, SCIM and retention rules point to Enterprise.

**4. Enable connectors read-only.** Turn on reading and search. Keep write actions blocked or on approval until the pilot shows how people use them.

**5. Build custom connectors for internal systems.** One MCP server per system, or a gateway in front of several, with the properties listed above.

**6. Pilot with one team.** Collect the questions people actually ask and where the answers were wrong or missing.

**7. Train people.** Teach what Claude can reach, when to check an answer against the source, and what not to paste. Engineering teams need separate, deeper enablement for Claude Code: project rules, permission settings, review workflow. That is what our [Claude Code training](https://goldenroz.com/claude-code-training/) covers, and [Claude Code best practices](https://goldenroz.com/blog/claude-code-best-practices/) is a starting point.

**8. Widen and review.** Add teams, add write actions one at a time, and review the logs.

## Common mistakes

- **Enabling every connector on day one.** Each one widens what a mistaken or manipulated request can reach.
- **A custom connector on a shared admin credential.** Every user inherits admin access through it.
- **Installing third-party MCP servers without review.** It is third-party code with access to your AI workspace.
- **Treating the plan as the project.** Seats are the small cost. The value comes from the connections and from people changing how they work.
- **No owner.** Someone must own connector settings, the MCP servers and the usage review.

## Where an independent firm fits

Anthropic sells the plans. It does not build the connection to your ERP, clean up your permissions or train your engineers on your codebase. That work is done by your own team or by a firm like ours. Goldenroz is independent, with no partner or reseller agreement with Anthropic. Our [Claude consulting](https://goldenroz.com/claude-consulting/) work covers use case selection, connector and permission design, custom MCP servers and team enablement. Clients hold their own Anthropic account and pay Anthropic directly.

If you are still deciding between vendors, see [Claude vs ChatGPT for business](https://goldenroz.com/blog/claude-vs-chatgpt-for-business/).

## What to do next

List the five systems your team would most want to ask questions of, and for each note whether a connector exists in Anthropic's directory and whether it holds personal or financial data. The ones without a connector are your MCP scope. Bring that list to a free one-hour consultation. See [Claude consulting](https://goldenroz.com/claude-consulting/) and [MCP server development](https://goldenroz.com/mcp-server-development/). Projects start at $25,000 at a fixed price per phase.

## Questions buyers ask

### What is Claude Enterprise?

Claude Enterprise is Anthropic's plan for larger organizations. As of November 2026, Anthropic's pricing page lists a 20-seat minimum and features beyond the Team plan, including SCIM user provisioning, audit logs, custom data retention controls and a Compliance API. Pricing is a per-seat fee plus usage.

### What are Claude connectors?

Connectors link Claude to apps and services so it can retrieve data and take actions there. Anthropic lists available connectors in a directory, and according to its documentation Claude inherits each person's permissions from the connected service. If a user cannot open a file in the source system, Claude cannot either.

### What is a Claude MCP server?

An MCP server is a program that exposes tools and data to AI applications through the Model Context Protocol. In Claude, a remote MCP server is added as a custom connector. This is how a company connects Claude to an internal system that has no connector in the directory.

### Does Anthropic train its models on business data?

Anthropic's privacy documentation states that by default it does not use inputs or outputs from its commercial products, including the Team and Enterprise plans and the API, to train models. The documented exception is data a user explicitly submits, such as feedback, which organization owners can turn off.

### Is Goldenroz an Anthropic partner?

No. Goldenroz is an independent engineering firm with no partnership or reseller agreement with Anthropic. We build MCP servers and Claude-based systems for clients, who hold their own Anthropic accounts and pay Anthropic directly.

---
Page: https://goldenroz.com/blog/claude-for-business/
Book a free one-hour call: https://calendly.com/daniarozin/daniel-rozin-consultation-1-hour-clone
