How to Govern Claude and Claude Code in the Enterprise

Teams are adopting AI assistants like Claude, and agentic coding tools like Claude Code. A real governance gap opens up. Employees use them to discuss sensitive business information, upload internal files, and generate code. All of it happens on a surface that traditional monitoring does not see [1]. Agentic coding tools add more to govern still: autonomous actions, access to tools and Model Context Protocol servers, and code execution. Security and compliance teams reasonably ask how they are supposed to govern any of it.

The good news is that this has become tractable. Anthropic’s Compliance API works alongside the broader Claude Enterprise controls. Together they turn Claude usage into something you can audit, monitor, and govern with the tools you already run. This guide covers what those controls do, and how to govern Claude Code specifically. It also covers where coverage is strong, where the gaps are, and how to implement it. The tooling is evolving quickly. Treat specifics as point-in-time, and verify against Anthropic’s official documentation.

The governance problem with AI assistants and coding tools

AI assistants are a new and hard-to-monitor surface. An employee can paste regulated data into a prompt: protected health information, cardholder data, privileged material, or regulated source code. They can upload confidential files and generate content or code. The DLP and SIEM tooling built for other applications does not necessarily catch any of it [2][3]. Agentic coding tools raise the stakes further. The tool is not just answering questions. It is taking actions, reaching into repositories and systems.

The blunt version is worth repeating to any leader who thinks this is handled. “Be careful what you paste” is not a policy [3]. Real governance needs visibility into what is happening, plus enforceable policy, audit trails, data protection, and access control. These are the same things you demand of any other enterprise system that touches sensitive data.

The Claude Compliance API

The core mechanism Anthropic provides is the Compliance API. It gives Claude Enterprise and Console customers programmatic access to their own Claude activity. That covers chats, files, and projects. It also covers Cowork and Claude Code sessions, plus the directory of users, roles, and groups [4]. Security, legal, and compliance teams use it to audit activity. They also retrieve or delete content, and feed events downstream [4].

In plain terms, it governs Claude usage itself [5]. It turns Claude into an auditable SaaS application. You get continuous monitoring, scriptable data-subject-access-request fulfillment, real-time policy alerts, and control evidence. That is what makes SOC 2, HIPAA, GDPR, and ISO 27001 obligations tractable [6].

It exposes data through a couple of access paths. An Activity Feed of per-event records is reachable with an Admin API key. Fuller access to conversation content and sessions needs a Compliance Access Key, created in the Claude admin settings [4][3]. It is also deliberately distinct from Anthropic’s analytics APIs. The Enterprise Analytics and Claude Code Analytics APIs return aggregated usage and cost figures for IT and finance teams. The Compliance API returns per-event records for security, legal, and compliance teams [4].

The reason this matters so much in practice is the integration ecosystem. In May 2026, Anthropic expanded the Compliance API with a large set of security and compliance integrations [5]. They span DLP, SIEM, identity, eDiscovery, and AI governance. Named partners include Microsoft Purview, CrowdStrike, Okta, Netskope, Zscaler, Wiz, Varonis, Snyk, and IBM Guardium, among others [5][8]. Say your security team already runs one of those platforms. Wiring Claude usage into your monitoring then takes hours, not weeks [3]. Most mid-market firms run at least one. This connects to the tooling covered in implementing AI governance with tools and Microsoft Purview. Purview is one of the integrations that can scan Claude conversation content against your existing classification rules.

Governing Claude Code specifically

Claude Code is an agentic coding tool, so it needs governance beyond chat monitoring, and Anthropic provides several layers.

Claude Code sessions are covered by the Compliance API. The same audit, retrieval, and policy machinery applies [4]. On top of that, Claude Code supports observability and control through OpenTelemetry. It streams metrics, events, traces, tool decisions, hook activity, and MCP activity to a collector you run. You can route that into your SIEM [6].

The agentic execution itself is covered too. Managed agents can run in self-hosted sandboxes and connect to private MCP servers [5]. Tool execution can therefore happen on your own systems. Your network policies, audit logging, and security tooling all apply to what the agent actually does. That last capability matters most for governing what an autonomous coding agent is allowed to touch.

The broader Claude Enterprise controls

The Compliance API is one part of a wider governance stack that Claude Enterprise provides [7]:

• Identity and access: SSO, domain capture, SCIM and just-in-time provisioning, and role-based access control. Admins govern exactly who can use which products and connectors.
• Data controls: custom data retention, customer-managed encryption keys, and US-only inference.
• Network controls: IP allowlisting, plus network-level control that blocks personal or non-corporate Claude instances from enterprise networks. A direct lever against shadow usage.
• Connector governance: admins approve which connectors are available and set per-tool permissions.
• Logging: audit logs, the Compliance API, and OpenTelemetry. All of them feed the DLP, SIEM, and monitoring tools your security team already runs.

Together these turn Claude from an ungoverned tool employees happen to use into a managed enterprise application. Access, data handling, and monitoring are all defined.

The honest gaps and caveats

This is genuinely strong, but it is not automatic and not complete. Using it well means knowing the limits.

It solves a large part of the problem, not all of it. One detailed practitioner assessment put it this way. The Compliance API is the most important thing Anthropic has shipped for enterprise governance. It addresses roughly half of the actual problem. It is strongest for Claude.ai chat and the Claude API [6].

Coverage varies by surface and is evolving. Different Claude surfaces joined the audit and compliance feed at different times. The guidance has been that regulated workloads should not run on a surface until it is fully covered [6]. This is actively changing. Confirm current per-surface coverage before putting regulated workloads on any newer surface. That applies to Claude Code and Cowork specifically.

It is an Enterprise-tier capability. The Compliance API requires Claude Enterprise or Console. A Primary Owner enables it, and parts of the specification sit behind an NDA review [6]. Teams on lower tiers do not have this governance layer at all. That is itself a reason regulated teams should be on Enterprise.

It governs Claude, not your whole AI estate. The Compliance API governs Claude usage specifically [3]. Governing every AI tool your organization uses requires broader coverage. That is exactly why the same security platforms integrate across many AI surfaces.

Treat the keys as sensitive. Compliance API keys are organization-scoped service-account credentials. Rotate them, store them in a vault, and audit access to the key store. At least one organization left one sitting in a chat message [6].

What implementing this looks like

A practical sequence, which is squarely the kind of governance-and-integration work worth getting right:

1. Write a real AI acceptable-use policy first. Name the approved tools. List the prohibited data classes, such as PHI, cardholder data, privileged material, and regulated source code. Define who approves new AI tools [3]. Policy before plumbing.
2. Get on the right tier. The Compliance API and the enterprise controls require Claude Enterprise or Console [6].
3. Turn on the access, network, and data controls. Configure SSO and role-based access for who can use what. Use network controls to block personal Claude instances from corporate networks. Set retention, encryption, and connector governance.
4. Wire activity into your existing security stack. Check whether your team already runs one of the integration partners, such as Purview, CrowdStrike, or Okta. Then stream Claude and Claude Code activity into your SIEM and DLP through the Compliance API. Add OpenTelemetry for Claude Code, with real-time policy alerts [3][6].
5. Use it for evidence and requests. Generate the audit evidence your frameworks require, and use the API for data-subject-access-request fulfillment and content retrieval or deletion.
6. Map the controls to your framework. Document which NIST AI RMF subcategories and EU AI Act requirements these controls satisfy. Governing Claude is then part of a coherent program rather than a standalone effort, as laid out in the governance pillar.

The bottom line

Claude and Claude Code are becoming part of how work and code get done. Governing that usage is a real requirement. Anthropic’s Compliance API, together with the Enterprise governance controls, makes it achievable. They turn Claude into an auditable, policy-governed surface. It plugs into the DLP, SIEM, and identity tools you already run.

It is not automatic, and not complete. It is an Enterprise-tier capability, coverage varies by surface, and it governs Claude specifically rather than your entire AI estate. The organizations that handle it well start with a real AI use policy. They turn on the access, network, and data controls, wire the activity into their existing security stack, then map it to their governance framework. Governing your AI tools has become as necessary as governing any other enterprise application, and, increasingly, just as doable.

---

Rolling out Claude or Claude Code and need to govern it? The path runs through a real AI use policy, the Enterprise access and network controls, and wiring the Compliance API into the security stack you already run.

---

Sources

1. Security Boulevard (2026). Anthropic Expands Claude Compliance API With 28 Enterprise Security Integrations (the monitoring gap for AI assistant usage; managed agents in self-hosted sandboxes and private MCP servers). https://securityboulevard.com/2026/05/anthropic-expands-claude-compliance-api-with-28-enterprise-security-integrations/
2. PurpleShield Security (2026). Claude’s New Compliance API: AI Governance Goes Mainstream (what the API exposes; the integration partners; the AI acceptable-use-policy guidance; that it governs Claude specifically). https://www.purpleshieldsecurity.com/post/claude-compliance-api-ai-governance
3. (See PurpleShield, above, on prohibited data classes, the integration shortcut, and metadata-versus-content.)
4. Anthropic. Compliance API — Claude Platform Docs (programmatic access to chats, files, projects, and Cowork and Claude Code sessions; access keys; the distinction from the analytics APIs). https://platform.claude.com/docs/en/manage-claude/compliance-api
5. (See Security Boulevard, above, on the 28 integrations and named partners including Microsoft Purview.)
6. General Analysis (2026). Claude Compliance API: Coverage, Gaps, and How to Use It (the “solves about half the problem” assessment; per-surface coverage caveats; Enterprise-only and NDA notes; key-management guidance; Claude Code OpenTelemetry observability). https://generalanalysis.com/guides/claude-compliance-api
7. Anthropic. Claude Enterprise Plan (identity and access, data controls, network controls, connector governance, audit logs, and the Compliance API). https://www.anthropic.com/product/enterprise
8. Anthropic Help Center. Get started with Claude Compliance API integrations (partner integrations such as Harmonic Security, IBM Guardium, Airia, and others governing Claude use across surfaces including Claude Code). https://support.claude.com/en/articles/15167101-get-started-with-claude-compliance-api-integrations