Claude Code and Claude Cowork Come to US Federal Agencies in FedRAMP High-Authorized Public Beta

Anthropic has opened a public beta of Claude Code and Claude Cowork inside Claude for Government Desktop, making both tools available to US federal agencies through a FedRAMP High authorized environment for the first time.

Before this launch, federal agencies that wanted to use Claude Code in a FedRAMP-authorized setup had to route through Amazon Bedrock or Google Vertex. Now, both Claude Code and Cowork are available directly in the Claude for Government Desktop client, on the same release cadence as commercial users.

What the two tools actually do

Claude Code is Anthropic’s agentic coding assistant. In a government context, it’s positioned at software modernization: the kind of work that involves understanding large legacy codebases, writing new services, reviewing pull requests, and generally accelerating the work of software teams building or maintaining systems that support public services.

Claude Cowork is a desktop-first assistant that works directly with files on your machine. Agency staff can hand off tasks like drafting memos, reviewing RFPs, working through casework, or building out decks. It launched commercially in January, and this federal rollout is part of a broader expansion that also brings Cowork to mobile and web for commercial Max subscribers.

The governance layer is the real story

The capability additions matter, but what federal agencies actually need to authorize and deploy AI is the governance scaffolding around it. That’s where this release spends most of its effort.

Tamper-evident audit logs. Every administrative action is recorded in a hash-chained log, meaning each entry cryptographically references the previous one. Retroactive tampering becomes detectable. Organization administrators can review these logs directly in the product, without needing to pull data out of the environment.

ATO-friendly usage exports. Usage exports contain metering data only. Agencies can respond to ATO and Inspector General requests without moving sensitive conversation content anywhere.

Department-level spend controls. Program offices can tie AI spend to appropriated funds using standard seats, or define custom seat tiers with their own spend limits and model restrictions. Usage is purchased in fixed increments with a hard not-to-exceed cap, and automatic burndown alerts fire before the balance runs low. Per-user and per-model tracking sits in the admin console.

SCIM group mappings. Administrators can map their existing identity provider groups directly to seat tiers, applying rate limits, dollar caps, and model restrictions without manual user management.

Layered sub-agency configuration. A department-level administrator can allocate seats and prepaid usage to sub-agencies while letting each manage its own users. Defaults cover what Claude can connect to, which features are available, and instructions that shape how Claude interacts with staff.

Two-person approval for sensitive operations. On Anthropic’s side, sensitive operations require two-person approval before they can proceed.

Data handling and where things run

Conversation history stays local on the agency-managed device. Inference transits to the FedRAMP High authorized cloud environment, but nothing is stored remotely by default. The FedRAMP authorization is held through Palantir Federal Cloud Service Supporting Services, with independent assessment by Schellman Compliance.

For IT teams handling deployment, the desktop client ships through standard MDM platforms, and administrators can control file mount-path allowlists via MDM settings. Anthropic is also publishing its FedRAMP Secure Configuration Guide as a public-facing document, and a penetration-test summary for the new desktop client is available through their trust center.

One practical note on the Cowork architecture: the agent loop runs outside a VM while code execution stays inside it. The VM enforces filesystem and network controls over anything the agent actually runs, while still allowing Claude to respond to the user during execution.

Procurement and pricing

There are two routes to procurement: directly through Anthropic’s sales team, or via Carahsoft for agencies that need a GSA Schedule vehicle. Anthropic remains the contracted and billing party in both cases, so agencies do not need to set up a separate cloud-provider relationship first.

Standard pricing is $60 per seat per month. There is currently a limited-time program making Claude for Government available at $1 per agency (unlimited seats) through August 2026, which is worth knowing if your agency is evaluating now.

What this means in practice

For authorizing officials and security teams, the combination of hash-chained audit logs, local data storage, a published secure configuration guide, and available pentest artifacts takes a meaningful chunk of ATO paperwork off the table.

For department administrators, the spend controls are genuinely useful: hard caps, burndown alerts, and per-model tracking address the practical problem of aligning AI usage to appropriated budget lines rather than discovering overruns after the fact.

For software teams, getting Claude Code in the same authorized environment where other agency work happens removes the friction of routing through a separate cloud service. Same model, same tooling, same cadence as commercial users.

For agency staff doing document-heavy work, Cowork’s file-native approach means RFP reviews, memos, and casework can happen without copying content into a web interface and hoping it stays contained.

The broader signal here is that frontier AI vendors are increasingly doing the compliance and governance work themselves rather than leaving it to system integrators. Whether that changes procurement patterns in federal IT is a longer story, but it does lower the activation energy for agencies that have been waiting on the authorization side of the equation to catch up.