Claude Code adds background MCP tasks and clearer permission failures
Anthropic’s November 6 release changes how long-running MCP calls complete and how unattended sessions report denied actions. Teams maintaining agent workflows have concrete compatibility checks to make.
Read the story
Claude Code v2.1.319, published on November 6, adds terminal support for MCP tasks. When a server runs a tool call as a task, Claude Code moves it into the background and delivers the result when it finishes. The release also provides CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND for workflows that need to wait inside the tool call. This behavior depends on the server using the task mechanism; it is not a claim that every existing MCP integration becomes asynchronous. A new permissions.denyInsteadOfPrompting setting denies calls that would otherwise wait for approval and lets the turn continue. It does not grant permission or establish that the requested action succeeded. Separately, Anthropic says claude -p now exits with status 1, rather than 0, when auto-mode actions are refused because no safety check is available. The notes also describe fixes for older MCP handshakes and for permission-check bypasses involving quoted Bash words, crafted environment prefixes and sed, plus a sandbox escape during worktree creation. These are the maintainer’s release claims; IDAKIM Digital has not independently tested the fixes. For auto mode behind gateways or with telemetry or feature flags disabled, the notes describe changed handling when the server cannot check a call. Review the exact release configuration guidance before rollout.
Why it matters
For teams running unattended jobs, test task completion delivery, denied-action handling and exit-code propagation in staging before updating managed environments. A continuing conversation can contain refused work, so verify the actual output artifact as well as the process status. Review gateway-specific settings with the operator instead of weakening permission checks to make a job finish.