Anthropic Enterprise-Managed Auth for MCP Connectors

Anthropic is trying to remove one of the last pieces of friction holding enterprises back from wiring Claude into their tools. On August 24, 2026, the company said enterprise-managed authorization for Model Context Protocol (MCP) connectors is now generally available, letting administrators provision third-party integrations for an entire organization through their identity provider instead of asking every employee to click through an OAuth screen.

The change targets a specific and stubborn problem. MCP lets AI applications call external tools and read data sources, but connecting each of those tools has historically meant per-user authorization — every employee individually granting Claude access to Slack, or Notion, or an internal database. For a company with thousands of seats and a security team that wants a say in what gets connected, that model does not scale. Anthropic’s answer is to move the decision up a level, to the admin.

What shipped

With enterprise-managed authorization, a Claude Team or Enterprise administrator can authorize a connector once, assign access through their existing identity-provider groups and roles, and have employees automatically receive the approved connection the first time they sign in to Claude. Anthropic describes the result as zero-touch OAuth: for end users, tools and data are connected automatically, with no individual authorization step.

The rollout begins with Okta as the supported identity provider. Administrators provision MCP connectors for their whole organization through Okta, mapping connector access to the groups and roles they already maintain. That means access decisions ride on the same directory a company uses for the rest of its software, rather than living in a separate, per-app permission sprawl.

Anthropic has implemented the capability in its shared MCP layer, so admins can authorize MCP servers for users across Claude, Claude Code, and Cowork from one place. A connector approved for a department flows to all three surfaces without separate setup for each.

The security angle admins asked for

The feature is as much about control as convenience. Alongside provisioning, administrators can enforce an identity-provider-only connection requirement for a given connector — a setting that prevents employees from linking their own personal accounts to a work integration.

That control addresses a quiet risk in the connector model. When authorization is left to individual users, nothing stops someone from connecting a personal Notion workspace or a personal Slack to a corporate Claude session, blurring the line between personal and enterprise data. By requiring that a connector be reached only through the company’s identity provider, an admin closes that path and keeps data from crossing between personal and enterprise environments.

The design also inherits the direction MCP itself has been moving. The protocol’s 2026-07-28 revision rewrote authorization around token validation and pushed the standard toward distributed, multi-tenant deployments — exactly the shape an enterprise needs when it is fanning one approval out to thousands of users. Enterprise-managed authorization is the productized form of that shift: centralized identity, delegated per-user access, and validation handled in the shared layer rather than by each connector on its own.

Which connectors are covered

The general-availability milestone came with an expanded roster. Anthropic added Datadog, Notion, and Slack to the list of connectors that support enterprise-managed authorization. They join a previously supported group that includes Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase.

The company said Exa, Miro, and Zoom are coming soon, signaling that the supported set is meant to keep growing rather than freeze at launch. The additions skew toward the tools knowledge workers live in day to day — observability, documentation, messaging, design, and project tracking — which is where connected AI agents tend to earn their keep first.

Who is already using it

Anthropic pointed to early enterprise deployments to make the case that the feature is production-ready rather than a preview. Organizations including HubSpot, Ramp, and Webflow are already using enterprise-managed authorization to roll connectors out across their teams.

That list matters because it maps to where Anthropic’s business has been heading. The company’s enterprise run-rate revenue reached roughly $47 billion by late May 2026, up from about $9 billion at the end of 2025, and more than 1,000 enterprise customers now spend at least $1 million a year. Those customers are precisely the ones for whom per-user OAuth was a blocker — large deployments where a security and IT function, not an individual employee, decides what Claude may touch.

The connector push also fits Anthropic’s broader enterprise strategy, which has increasingly emphasized the workflow and control layers around models rather than raw model capability alone. As rivals ship comparable frontier models, the differentiator shifts to whether an organization can trust an AI system with its data, govern what it accesses, and manage that access at scale. Enterprise-managed authorization is a bet on that control layer.

How it differs from the old flow

Under the previous model, connecting a tool to Claude followed the familiar consumer OAuth pattern: a user clicks “connect,” gets redirected to the tool’s login, grants scopes, and returns with a token tied to their individual account. Multiply that across every tool and every employee, and an enterprise ends up with thousands of independent grants that no one centrally tracks, revokes, or audits.

Enterprise-managed authorization inverts the flow. The admin authorizes the connector once and binds it to identity-provider groups; the user never sees an OAuth prompt. Access is granted, changed, or revoked by adjusting group membership in the identity provider — the same lever IT already pulls to on-board and off-board staff. When an employee leaves and is removed from the directory, their connector access goes with them, without anyone hunting down individual tokens.

That is a meaningful change for anyone who has thought about how OAuth tokens differ from API keys in a large deployment, or who has wrestled with provisioning through standards like SCIM. It brings MCP connectors in line with how the rest of enterprise software handles access: centrally, through the directory, on the admin’s terms.

What it means

For enterprises, this lowers the cost of saying yes to connected AI. The friction of per-user OAuth was never just annoyance — it was a governance gap that security teams could not sign off on. By routing connector access through the identity provider, Anthropic lets a CISO approve Claude’s reach once and manage it with the tools they already trust. That removes a common reason enterprise pilots stall before they scale.

For Anthropic, the feature is a competitive move dressed as plumbing. When frontier models converge in capability, the company that makes AI easiest to deploy safely inside a large organization wins the seats. Enterprise-managed authorization, shared across Claude, Claude Code, and Cowork, is designed to make Claude the default agent layer for companies that have already standardized on an identity provider — and to make switching away costlier once connectors are provisioned org-wide.

For the MCP ecosystem, it is a signal that the protocol has crossed from developer curiosity into enterprise infrastructure. Vendors that build MCP servers — the Datadogs, Notions, and Slacks of the list — now have a reason to support enterprise authorization patterns as a first-class feature, because that is what unlocks large-account adoption. Expect competitors to match the admin-managed model quickly; centralized, identity-provider-based authorization is fast becoming table stakes for any serious agent connector framework.

The open question is breadth. Launching with Okta covers a large share of enterprises but not all of them, and the supported-connector list, while growing, is still a fraction of the tools a big company runs. How fast Anthropic adds identity providers and connectors — and how cleanly the model handles the messy realities of overlapping groups and least-privilege access — will determine whether this becomes the standard way enterprises connect AI, or just a strong start.