Anthropic's Claude is now living inside Notion. The company quietly rolled out Claude agents as a beta feature this week, letting Business and Enterprise plan users spin up AI assistants that work directly within their Notion workspace. Unlike the existing Custom Agents built entirely in Notion, these new agents run on Anthropic's infrastructure while presenting an interface native to Notion—giving teams Claude's coding chops and file-editing capabilities without leaving the app they already live in.

Setup and Access Requirements

Getting started requires navigating to Agents in Notion's sidebar, clicking New Agent, and selecting Claude. Users can choose from default templates (some with GitHub integration via personal access tokens) or build an agent from scratch with custom instructions and triggers. Notably, no separate Anthropic account is needed—usage flows through Notion credits, charged per run rather than by token count. The feature is rolling out gradually, so some workspaces won't see it immediately.

Permissions Model Differs From Standard AI Features

Claude agents in Notion operate on a permission structure that breaks from typical workspace inheritance rules. Access is set at the agent level and does not cascade from whoever initiates a run. Admins can grant different groups varying access: can view and interact, can edit, or full access to content. For Enterprise and HIPAA workspaces, Claude agents are disabled by default—workspace owners must explicitly enable them in Settings → Notion AI → Agent → Manage external agents before anyone can use them.

Data Retention Caveats Worth Noting

Here's where things get interesting for security-conscious teams: Notion's zero data retention commitments do not apply to Claude agents. The feature runs via Claude Managed Agents (CMA), which is stateful by design—Anthropic stores session data between requests to maintain context across interactions, making ZDR incompatible with how these agents function. Organizations operating under strict data handling requirements should factor this in before enabling the feature.

Feature Set Enables Real Workflow Delegation

Claude agents can chat with team members, work from shared docs and task boards (assigning a task and asking for help is explicitly supported), create or update content when granted edit access, browse the web, and even call other agents mid-session. They see only what's been explicitly shared with them—if an agent can't access something, it typically means sharing settings need adjustment rather than a system bug.

How This Differs From MCP

Notion's documentation is explicit: this is not the same as using Claude with Model Context Protocol. With MCP, you work in the Claude app and pull data from Notion. With Claude agents in Notion, the entire workflow stays inside Notion—you never leave to use a separate application. This represents a fundamentally different integration approach aimed at keeping AI assistance embedded within existing workflows rather than bridging between tools.

Key Takeaways

  • Beta access requires Business or Enterprise plans; rollout is gradual
  • No Anthropic account needed—billing runs through Notion credits per run
  • Permissions are agent-level, not inherited from the user triggering a task
  • Zero data retention does NOT apply—CMA stores session state between requests
  • Different from Custom Agents (built in Notion) and different from MCP integrations

The Bottom Line

Notion's embedding of Claude agents signals a deeper commitment to making AI a first-class participant in knowledge work rather than a peripheral tool you context-switch to. But the CMA data retention issue is a real constraint for regulated industries—no amount of UI polish erases that these agents maintain state outside Notion's usual privacy commitments. Teams should evaluate whether their compliance requirements can accommodate that trade-off before diving in.