39 Bringing Claude Code and Claude Cowork to Government
39.1 Overview
On July 7, 2026, Anthropic announced the availability of Claude Code and Claude Cowork in Claude for Government Desktop, extending the company’s government-grade offering to include agentic coding and collaborative work capabilities. This release brought the two most powerful Claude surfaces into a FedRAMP High authorized environment, giving federal agencies the tools to adopt agentic workflows within the strict security and compliance frameworks that govern government IT.
The announcement was significant for several reasons. Government agencies had been eager to adopt AI coding tools — the productivity gains demonstrated in the private sector were directly applicable to the sprawling legacy codebases that many agencies maintained. But the security requirements for federal systems are far more demanding than those for commercial enterprises. Static API keys, cloud-dependent authentication, and commercial billing models are often non-starters in government contexts.
Claude for Government Desktop addressed these challenges with a purpose-built environment that combined the power of Claude Code and Cowork with the security, auditability, and billing flexibility that federal procurement and compliance frameworks demand.
39.3 Features for Government
39.3.1 Configuration Defaults
Government deployments required secure-by-default configurations. Claude for Government Desktop shipped with a set of configuration defaults that aligned with federal security requirements:
- Conservative model defaults: Defaulting to models and settings that balanced capability with predictability.
- Restricted tool use: Limiting file system access, network operations, and tool execution to approved patterns.
- Data handling controls: Ensuring that conversation data, code, and file contents were handled according to federal data classification guidelines.
# Example government configuration defaults
security:
data_classification: "CUI" # Controlled Unclassified Information
conversation_retention: 0 # No cloud retention
telemetry_to_vendor: false # No data sent to Anthropic
file_access:
mode: "allowlist"
approved_paths:
- "/home/user/projects/*"
- "/opt/gov/shared/*"
network:
external_access: false
approved_endpoints:
- "https://claude.gov.internal"39.3.2 Spending Controls
Government budgeting operates under appropriations — funds allocated by Congress for specific purposes, with specific time periods and spending limits. Commercial SaaS billing models (pay-per-use, overage charges) don’t map cleanly to this framework.
Claude for Government addressed this with spending controls designed to fit appropriation structures:
| Feature | How It Works | Government Value |
|---|---|---|
| Fixed increments | Purchase usage in pre-defined blocks | Aligns with obligated funds |
| Not-to-exceed cap | Hard ceiling on total spend | Prevents anti-deficiency act violations |
| Period-based limits | Spending bounded by fiscal year/quarter | Matches appropriation periods |
| Department-level allocation | Budgets per department/sub-agency | Maps to organizational structure |
The “not-to-exceed cap” was particularly important for government customers. Under the Antideficiency Act, federal agencies cannot obligate more funds than Congress has appropriated. A hard spending ceiling ensured that Claude usage could never inadvertently exceed available appropriations — a legal requirement, not just a budgeting preference.
39.3.3 Tamper-Evident Audit Logs
For government systems, auditability is not optional — it is a regulatory requirement. Claude for Government introduced tamper-evident audit logs using a hash-chaining mechanism:
Log Entry 1: hash = SHA256(entry_1_data || previous_hash || seed)
│
▼
Log Entry 2: hash = SHA256(entry_2_data || hash_of_entry_1 || seed)
│
▼
Log Entry 3: hash = SHA256(entry_3_data || hash_of_entry_2 || seed)
│
▼
...Each log entry’s hash incorporated the hash of the previous entry, creating a chain that could not be broken without detection. If any entry was modified or deleted, the hash chain would break, and the tampering would be immediately visible during audit.
This design meant:
- Unauthorized modifications were detectable.
- Deletions left gaps in the chain.
- Insertion of fake entries required recomputing all subsequent hashes.
- Verification was possible by recomputing the chain from the beginning.
# Simplified hash-chain verification
import hashlib
def verify_audit_chain(entries: list[dict], seed: bytes) -> bool:
"""Verify the integrity of a hash-chained audit log."""
previous_hash = seed
for entry in entries:
expected_hash = hashlib.sha256(
json.dumps(entry['data'], sort_keys=True).encode() +
previous_hash +
seed
).hexdigest()
if entry['hash'] != expected_hash:
return False # Tampering detected
previous_hash = entry['hash']
return True # Chain is intact39.3.4 Two-Person Approval for Sensitive Operations
Certain operations in government systems require two-person control — a security principle where sensitive actions must be authorized by two separate individuals. Claude for Government implemented this for operations like:
- Granting elevated access to another user.
- Modifying security policies.
- Exporting large datasets from the system.
- Disabling audit logging (which should never happen in production).
When a two-person-controlled operation was initiated, the request was queued for approval by a second authorized individual. Only after both approvals were recorded did the operation execute — and both approvals were captured in the hash-chained audit log.
39.3.5 ATO Support
Authority to Operate (ATO) is the formal authorization that a federal agency must grant before a system can process government data. Obtaining an ATO requires extensive documentation, security assessments, and risk analysis.
To accelerate ATO processes for agencies adopting Claude, Anthropic provided:
- Pre-completed security documentation aligned with NIST SP 800-37 (the RMF framework).
- System Security Plan (SSP) templates pre-populated with Claude for Government’s control implementations.
- Penetration test results from independent third-party assessors.
- Continuous monitoring artifacts demonstrating ongoing compliance.
Anthropic published a FedRAMP Secure Configuration Guide that documented every security-relevant configuration, the rationale behind each default, and step-by-step instructions for agency security teams to validate the configuration in their own environment. This transparency was essential for agencies that needed to independently verify security claims rather than taking them on faith.
39.4 Organizational Structure
39.4.1 Department-Level Administration
Government IT often follows a hierarchical structure: a department (e.g., Department of Defense, Department of Veterans Affairs) contains multiple sub-agencies (e.g., Army, Navy, Air Force within DoD), each with their own IT governance.
Claude for Government supported this structure with department-level administration:
- A department administrator could manage the overall deployment.
- Sub-agency administrators could manage their own users and policies within bounds set by the department admin.
- Spending limits could be allocated from the department down to sub-agencies.
- Policies could be inherited (department-level defaults) or overridden (sub-agency-specific).
Department of Example (DoE)
├── Department Admin: Alice
├── Budget: $500,000/year
│
├── Sub-Agency: Bureau of IT
│ ├── Sub-Agency Admin: Bob
│ ├── Allocated Budget: $200,000/year
│ └── Users: 150 developers
│
├── Sub-Agency: Bureau of Research
│ ├── Sub-Agency Admin: Carol
│ ├── Allocated Budget: $150,000/year
│ └── Users: 80 researchers
│
└── Sub-Agency: Bureau of Operations
├── Sub-Agency Admin: Dave
├── Allocated Budget: $150,000/year
└── Users: 120 analysts39.4.2 SCIM Group Mappings
System for Cross-domain Identity Management (SCIM) was supported for automated user provisioning and group management. Government identity providers (like PIV/CAC-based systems) could use SCIM to:
- Automatically create Claude accounts when users were added to an IdP group.
- Automatically remove access when users left or changed roles.
- Map IdP groups to Claude roles and permissions.
// SCIM group mapping example
{
"displayName": "DoE-Developers-Claude-Code",
"members": [
{"value": "user-001", "display": "Alice Chen"},
{"value": "user-002", "display": "Bob Martinez"}
],
"meta": {
"resourceType": "Group",
"claude_role": "developer",
"claude_service": "code"
}
}This ensured that access management was driven by the agency’s authoritative identity source, not by manual account creation in Claude.
39.5 Penetration Test Summary
To build trust with agency security teams, Anthropic made a penetration test summary available. This document, produced by an independent third-party security firm, summarized:
- Scope: What systems and interfaces were tested.
- Methodology: The testing approach (OWASP, PTES, etc.).
- Findings: Vulnerabilities discovered, categorized by severity.
- Remediation: How identified issues were addressed.
- Residual risk: Any remaining risk after remediation.
Penetration test summaries are carefully scoped documents — they provide enough detail for security teams to assess risk without exposing specific vulnerabilities that could be exploited. Anthropic’s willingness to share this level of detail was a deliberate transparency choice for the government market.
39.6 Use Cases in Government
39.6.1 Legacy Code Modernization
Many federal agencies maintain codebases written decades ago — COBOL, Fortran, early Java. Claude Code’s ability to understand, document, and migrate legacy code made it particularly valuable for:
- Understanding legacy systems: Having Claude read and explain undocumented code.
- Generating documentation: Creating technical documentation for systems that had none.
- Migration planning: Analyzing what would need to change to modernize.
- Test generation: Creating test suites for code that had never been tested.
39.6.2 Cybersecurity Operations
Claude Cowork and Code could support cybersecurity teams in:
- Analyzing malware and generating detection signatures.
- Reviewing security logs for anomalous patterns.
- Drafting incident response plans and reports.
- Generating compliance documentation for security controls.
39.6.3 Research and Analysis
Government research agencies could use Claude for:
- Literature reviews: Summarizing large bodies of research.
- Data analysis: Processing and analyzing datasets.
- Report drafting: Generating first drafts of technical reports.
39.7 Key Takeaways
- Claude Code and Cowork are now available in a FedRAMP High environment. The full agentic coding and collaborative work experience, within the highest federal security authorization.
- Configuration defaults are secure-by-design. Conservative model defaults, restricted tool use, and federal data classification controls built in from the start.
- Spending controls fit government appropriations. Fixed increments, not-to-exceed caps, and period-based limits align with how federal funding actually works — preventing Antideficiency Act violations.
- Tamper-evident audit logs use hash-chaining. Any modification, deletion, or insertion of log entries is cryptographically detectable, meeting federal audit requirements.
- Two-person approval controls sensitive operations. Elevated access, policy changes, and data exports require dual authorization, with both approvals recorded in the audit chain.
- ATO support accelerates adoption. Pre-completed security documentation, SSP templates, and published penetration test summaries reduce the time to obtain an Authority to Operate.
- Department-level administration mirrors government structure. Department admins can delegate to sub-agency admins with budget allocation and policy inheritance.
- SCIM integration automates identity management. Government IdP groups map directly to Claude roles, with automatic provisioning and deprovisioning.
- The FedRAMP Secure Configuration Guide and penetration test summary provide transparency. Agency security teams can independently verify security claims rather than taking them on faith.
- Claude for Government is purpose-built, not retrofitted. Every feature — from billing to audit logs to administration — is designed around the realities of federal IT governance.