Claude Code v2.1.285 Adds Provider Restrictions, Desktop Integration, and Subagent Reliability Fixes
What changed
Anthropic released Claude Code v2.1.285, focusing on stricter environment controls, smoother desktop workflows, and deeper fixes for background subagents and remote sessions.
New capabilities include CLAUDE_CODE_DISABLE_WEB_FETCH to completely block the WebFetch tool, and CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES to cap automatic re-sends when a non-streaming fallback times out. A new claude --desktop command launches the desktop application targeting the current working directory or resumes an existing session using --continue or --resume <id>. Plugin management received two upgrades: claude plugin configure <plugin> now displays missing options and accepts values via --values-stdin, while claude plugin install --config supports <server>.<key>=<value> syntax to pre-configure bundled .mcpb MCP servers without navigating the interactive menu. Additionally, the allowedProviders managed setting lets administrators restrict which endpoints a machine can target, including Anthropic, Bedrock, Vertex AI, Foundry, Mantle, custom endpoints, and AWS gateways.
The release also patches several stability and correctness issues. Background subagents created with CLAUDE_CODE_FORK_SUBAGENT=1 now properly inherit their parent’s permission modes and run agent calls in the foreground. Plugin installations over SSH correctly respect GIT_SSH and core.sshCommand, and the installer refuses packages that differ only by punctuation or casing to prevent cache collisions. Remote Control gained retry logic for failed file attachments, corrected message read-state tracking, and ensures queued messages survive terminal closures. Other fixes address redacted log leaks, sandbox auto-approval false positives on inline scripts, mid-session MCP server cleanup, and a corrected help string for claude remote-control --help that previously misstated the default behavior for the --[no-]chrome flag.
Why it matters
These changes shift Claude Code toward more predictable enterprise and automated workflows. Restricting API providers and disabling web fetching reduces blast radius in restricted networks. The desktop CLI flag and improved session resumption make local-to-desktop handoffs seamless. Subagent reliability improvements—especially around permission inheritance and foreground execution—prevent silent failures in complex orchestration chains. Finally, tightening plugin installation validation and fixing Remote Control state tracking reduces friction for teams managing shared environments or headless CI/CD pipelines.
What to do next
Update your CLI to v2.1.285 and audit any allowedProviders configurations if you manage multi-account deployments. Test CLAUDE_CODE_DISABLE_WEB_FETCH in isolated environments before applying it to production automation runners. If you rely on background subagents, verify that CLAUDE_CODE_FORK_SUBAGENT=1 now returns expected results without manual intervention. For plugin-heavy setups, use the new --config syntax to bake MCP server credentials into installation scripts, eliminating manual configuration steps. Note that recent releases have continued refining auto-mode and agent caching behaviors, so reviewing Claude Code v2.1.280 Introduces Opus 5.5, Tightens Auto-Mode, and Fixes Agent Caching may provide useful context for ongoing pipeline adjustments.
Caveats: Sessions will still halt if they encounter unparseable managed settings files or general read failures; only OS-level denials are bypassed with a warning. Cloud sessions and /remote-env reads remain capped at retrieving only the twenty most recent environments per account. Non-streaming timeout retries are strictly limited by the new environment variable, which may slow recovery during prolonged network instability.
Sources:
Sources
Share Article
If this article helped you, please share it with others!