Claude Team Admin Console: Setting Policy by Data Governance

1. Admin Console — main structure

When you start a Team plan (Standard/Premium) you get a Workspace. The owner enters Settings to see these main tabs:

• Members — list of employees in the plan, seat type (Standard/Premium), role (Owner/Admin/Member)
• Billing — invoices, renewal, seat count used/available
• Usage Analytics — message volume, Claude Code usage, model distribution (aggregate and per-user)
• Connectors — enable/disable Google Drive, Slack, GitHub, Microsoft 365 at the workspace level
• Skills — Skills the admin deploys for the workspace to use together
• Tool Permissioning (Premium) — what tools Claude Code is allowed to call, what's banned
• MCP Configuration (Premium) — list of MCP servers the workspace allows
• Audit Log — records of admin actions (e.g., user added, seat changed, connector approved)
• Data Settings — workspace-level privacy/retention controls

2. User & Seat Management — add/inactivate/switch seats

The heart of governance is controlling who can access the workspace. The first place admins use most is Settings → Members.

• Add a user: Invite by email — invitees must accept within 7 days (per Anthropic T&C)
• Assign roles:
  • Owner — full access, including billing and workspace deletion
  • Admin — manage members, settings, view analytics — cannot delete the workspace
  • Member — regular user, no admin settings
• Change seat type: Standard ↔ Premium can be switched anytime (cost is prorated per Anthropic T&C)
• Inactivate a user: "Remove from workspace" — the user loses access; workspace chat history stays in the org (it doesn't follow the user)
• Transfer ownership: if the original owner leaves, ownership can be transferred to someone else

3. Data Settings & Workspace-Level Policy

The section that maps directly to Data Governance:

3.1 No Training on Team Data

Anthropic states clearly in its documentation that Team/Enterprise content is not used to train models — this is the default with no opt-in needed, unlike Free/Pro/personal Max where users toggle this themselves (Anthropic adjusted the Personal default in 2025).

• Member chat messages are not fed into training
• Files uploaded to Projects/Artifacts are not used to train
• Claude Code transcripts are not used to train
• Messages flagged for trust & safety review may be accessed by Anthropic under its abuse policy — separate from training data

3.2 Data Retention

• Default: member chats are kept in the workspace as long as the user wants them
• Members can delete their own chats — Anthropic removes them from the backend within its retention window (typically 30 days)
• Admins cannot directly fetch a member's files or chats — that requires a support request or data export API (where available)

3.3 Compliance Posture

• SOC 2 Type II — applies to Team/Enterprise
• BAA (HIPAA) — Enterprise can sign a Business Associate Agreement
• DPA (Data Processing Agreement) — Anthropic provides DPA templates for EU/UK and PDPA-equivalent regions
• GDPR/PDPA: Team plan creates a contractual data-processor relationship between the company and Anthropic — different from Personal, where the user is their own data subject

4. Privacy Boundary — Can Admins Read Member Chats?

This is the most important question of the article — and the short answer is by default, no.

What admins see vs. don't see

Item	Admin sees?	Notes
Member list, role, seat type	✓	Full — used for provisioning decisions
Member message/prompt count	✓ aggregate	Numbers, not content
Claude Code usage (time, tokens, command count)	✓	Drives the spend cap
Connectors a member has enabled	✓	Because the admin controls the allowlist
Member chat content via UI	✗	Admin Console has no "read another's chat" button
Member chat via Data Export	⚠ Primary Owner can	Export includes chats, uploaded files, usage patterns
Content in a member's private Project	✗	Member must share to the workspace first
Files uploaded in member chats	✗	Admin only sees the event, not the file
Audit log of other admins' actions	✓	Owner/Admin can view

What Primary Owner Data Export Actually Covers

Anthropic states in its Help Center that the Primary Owner of Claude for Work (Team/Enterprise) can request a data export of the workspace, which includes "conversations with Claude, uploaded files, and usage patterns" of members.

• Available to Primary Owner only (not every Admin)
• Comes as a file — not a "read user's chat in real time" capability
• Team plan audit logs do not export chat titles or content (only unique identifiers) — but direct data export does include content
• Enterprise plan adds a Compliance API + eDiscovery for legal hold, which Team plan does not match — suitable for regulated organizations

From an HR/employee angle — communicate clearly that a workspace account is not a private mailbox; it's a company asset the org can access under policy.

Anthropic designs Claude Team so admins see metadata about employees, not content — a clear separation from MDM/EDR systems that can read employees' screens. For IT managers it means — you control access, but you can't control the content of users' conversations with AI directly. That has to come from policy and user training.

- Anthropic's privacy-preserving governance design

5. Sharing — Projects, Prompts, Skills in a Workspace

One of Team's advantages over Personal: share tools across the workspace.

5.1 Shared Projects

• Members create a Project and choose visibility: Personal, Workspace, or Specific members
• Workspace Projects = everyone has access — good for "HR knowledge base," "Onboarding guide," "Customer FAQ"
• Owners/Admins can archive workspace-level Projects

5.2 Shared Skills

• Skills = capabilities a user can create (e.g., "Write a proposal in our company format," "Analyze P&L for Saeree")
• Personal: Skills are tied to the user's account
• Team: Admins deploy Skills at the workspace level — every member uses the same Skill, like a "company-wide preset"
• Example: HR creates a Skill "Screen resumes against our criteria" → admin deploys → every recruiter uses it the same way

5.3 Shared Prompt Templates

• Team plan supports a workspace-level prompt library
• Reduces prompt drift — the team doesn't need to recreate prompts every time
• Good for repeated patterns: "Draft an email in our company tone," "Summarize meetings in our template"

6. MCP & Connector Governance — What Can You Restrict?

MCP (Model Context Protocol) and connectors are where Personal and Team differ most — because they're the "gateway" for AI into your org's systems.

6.1 Connectors (Google Drive, Slack, GitHub, M365)

Connector control	Personal (Pro/Max)	Team (Std/Premium)
Who enables/disables connectors	The user	Admin controls the allowlist
See the scope the connector requests	User confirms	Admin pre-approves
Block specific connectors	✗	✓
Audit who connected what, when	✗	✓

6.2 MCP Server Allowlist (Premium)

MCP is an open standard for Claude to connect to any system via a server — the trade-off is MCP servers come from many vendors and not all are trustworthy.

• Personal: the user can add any MCP server — including ones the org's admin hasn't vetted
• Team Premium: admins set an allowlist of MCP servers — members can only add ones on the list
• Example: a company allows MCPs only for internal Postgres, GitHub Enterprise, and Linear — blocks third-party connectors that haven't passed security review

6.3 Tool Permissioning for Claude Code (Premium)

Claude Code is an agent that runs commands in a terminal — admins control what it cannot do:

• Block rm -rf, sudo, system-file deletion
• Block edits outside the project directory
• Restrict access to specific environment variables
• Block command groups (network egress, package installation)

7. Spend Cap & Cost Controls (Premium)

Claude Code is the feature that can "burn dollars" fastest if uncontrolled — Premium has guardrails:

• Org-level cap: workspace exceeds $X/month → automatic block
• Per-user cap: each member has a budget — once exceeded, Claude Code stops new agent runs for that user
• Alert thresholds: notifications at 50% / 75% / 90% of budget
• Analytics breakdown: who's burning the most tokens, which projects are heaviest

8. Personal vs Team — What Personal Cannot Do

Capability	Personal (Pro/Max)	Team (Standard/Premium)
Admin Console	✗	✓
User provisioning/roles	✗	✓ Owner/Admin/Member
Audit log	✗	✓
Shared Projects	✗	✓ Workspace-wide
Shared Skills (admin-deployed)	✗ user-only	✓ Workspace Skill
Connector allowlist	✗ user enables	✓ admin-controlled
MCP server allowlist	✗ user adds any	✓ (Premium)
Claude Code tool permissioning	✗ user-managed	✓ (Premium)
Per-user spend cap	✗	✓ (Premium)
No training on data (default)	Opt-in	✓ default
Central billing (one invoice)	✗ per-user invoices	✓
Offboarding (employee leaves)	Admin can't	Remove seat immediately

9. Skill vs MCP — Are They Different on Personal vs Team?

Users ask this often — Skill and MCP are two things Claude users frequently confuse. Here's the breakdown:

9.1 Skill vs MCP — Side-by-Side

Aspect	Skill	MCP (Connector)
What it is	A capability/preset for how to work — e.g., "Write in our company style"	A channel for Claude to access external data/services — e.g., Google Drive, GitHub, internal APIs
What it emphasizes	Thinking patterns / work formats	Data access and triggering external actions
Examples	"Analyze P&L," "Screen resumes," "Write press releases"	"Fetch a doc from Drive," "Read Gmail," "Open a GitHub issue"
Can Personal use it?	✓ Users can create their own	✓ Users can connect any MCP server
Team — what admins control	Deploy Skills workspace-wide — everyone uses the same version; admin updates once	Set the allowlist — enable/disable connectors and MCP servers as policy
Governance risk	Low — Skills don't reach outside	High — MCPs both read and send data outward

9.2 Real-world examples

• HR Skill: "Screen resumes against our 5 criteria" — admin deploys → every recruiter uses the same version; no drift
• Finance Skill: "Analyze budget variance" — fixed structure, reduces time-to-analysis
• Sales MCP: connect to internal CRM → Claude answers customer questions
• Dev MCP: connect to internal Postgres + GitHub Enterprise → Claude Code can read schemas and repos

10. Initial Setup Workflow — Your First 30 Minutes With a Team Plan

1. Set the workspace name + upload a logo — Settings → Workspace → Brand
2. Add a backup Owner — so you're not locked out if the primary owner leaves
3. Invite members by team — Settings → Members → Invite
4. Assign seat types by role — devs = Premium; marketing/sales/admin = Standard
5. Enable Audit Log + Data Settings — confirm the default "No training on data" is on (default)
6. Set the connector allowlist — enable only Google Drive and GitHub for the company; disable external connectors you don't use
7. Set the MCP allowlist (Premium) — add internal MCP servers that have passed security review
8. Set the Claude Code spend cap (Premium) — start at $20-50/user/month; adjust upward based on usage
9. Deploy 1-2 starter Skills — e.g., "Write email in our company tone"
10. Communicate with the team — send an internal policy doc on what may and may not be put into Claude

The Answers — Direct Replies to What You Asked at the Top

1. How do admins configure policy by our org's data governance? → Settings → Data Settings (retention/training), Members (access), Connectors + MCP + Tool Permissioning (scope of work)
2. What do admins see? → Member list, role, seat, aggregate usage analytics, Claude Code metrics, audit log of admin actions, billing
3. Can admins read member chat logs? → Not in the Admin Console (no UI to open another's chat) — but the Primary Owner can request a workspace data export that includes conversations, uploaded files, and usage patterns, per Anthropic's Help Center
4. What can be shared? → Projects (workspace-wide), Skills (admin-deployed), Prompt templates, MCP/Connectors on the allowlist
5. What can be restricted? → External connectors, MCP servers not in the allowlist, dangerous tools for Claude Code, members exceeding their spend cap
6. What can Personal not do? → Admin Console, audit log, shared Skills/Projects, connector allowlist, MCP allowlist, tool permissioning, spend cap, central billing — Personal has none of these
7. Skill vs MCP — different on Personal vs Team? → Skills on Personal stay with the user; Team admins deploy them workspace-wide. MCP on Personal lets users add any server; Team Premium admins set the allowlist.

References

• Claude Help Center — What is the Team plan?
• Anthropic — Claude Code and new admin controls for business plans
• Anthropic Trust Center — SOC 2, BAA, DPA, training data policies
• Anthropic Privacy Policy
• Model Context Protocol — Open Standard
• Personal Data Protection Committee (PDPC), Thailand