Image: Anthropic / anthropic.com Claude Code 2.1.248: Restricted Mode, Enterprise Credits, and Cross-Session Messaging Across Cloud Platforms
The August 28 release adds a --restricted flag for sandboxed deployments, a /usage-credits command for enterprise limit increases, cross-session messaging on Bedrock and Vertex, and a fix for a recurring prompt-cache miss.
Claude Code 2.1.248 shipped on August 28 with four changes worth noting: a new restricted execution mode, enterprise usage credits, expanded cross-session messaging, and a fix for a prompt-cache miss that had been firing roughly once an hour in long sessions.
Restricted mode
The new --restricted flag strips out Claude Code’s built-in command and code execution tools and limits file access to the working directory. It’s aimed at deployment scenarios where you want the model to read and edit files but not run arbitrary shell commands or reach outside a defined scope.
claude --restrictedThis fills a gap for teams integrating Claude Code into CI pipelines or sandboxed review environments where full tool access is too permissive. The flag removes the execution surface without requiring custom configuration files or permission overrides.
Enterprise usage credits
Enterprise organizations can now request higher usage limits directly from within Claude Code using the new /usage-credits command. Previously, teams hitting plan ceilings had to go through Anthropic’s support channels or the Claude Console. This brings the request flow into the tool itself.
Cross-session messaging on Bedrock, Vertex, and Azure Foundry
The SendMessage and ListAgents tools, which let Claude Code sessions discover and communicate with each other, now work on AWS Bedrock, Google Vertex, and Azure AI Foundry. Previously, cross-session messaging was limited to direct Anthropic API connections. Teams running Claude Code on cloud provider deployments can now use these coordination primitives the same way they would on a standard setup.
Prompt-cache miss fix
The release addresses a bug that caused a prompt-cache miss roughly once per hour in long sessions. The root cause: tool definitions were being re-rendered after an OAuth token refresh, which broke the cache hit. Long-running sessions on the Anthropic API or through operator setups with OAuth should see better token efficiency after this fix.
Other fixes in this release: sessions no longer disappear from the desktop app after 30 days, login screen behavior is more consistent across platforms, and MCP server authentication handling was improved.
Full changelog: anthropic.com