What happened
Anthropic’s Claude Code release notes list version 2.1.319 as the latest release on the page. The release adds support for MCP tasks in the terminal. When an MCP server runs a tool call as a task, Claude Code now moves it into the background immediately and tells Claude the result when the task finishes. Teams that want the tool call to wait synchronously can set CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND.
The release also changes auto mode when telemetry or feature flags are off, including when Claude Code is behind an LLM gateway. Calls that the server cannot check are now presented for confirmation or denied; the previous fallback to the built-in classifier is removed. Anthropic says Bedrock, Vertex and Foundry are unchanged. For those who want to keep using the built-in classifier in the changed cases, the notes specify CLAUDE_CODE_AUTO_MODE_SERVER=0. A second setting, CLAUDE_CODE_AUTO_MODE_SERVER=1, enables the server-side check behind an LLM gateway even when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set. If managed settings enable that latter variable, they must also set the server-mode variable to 1.
Other additions affect permissions and operational controls. The new permissions.denyInsteadOfPrompting setting denies a tool call that would otherwise wait for a permission prompt, allowing the turn to continue; the release says it applies in terminal sessions, subagents and teammates. Developers can also use /claude-api latency-optimize to identify waiting time in Claude API code and try faster settings against an evaluation. A new environment variable, CLAUDE_CODE_DISABLE_FEATURE_FLAGS, stops feature-flag fetching without disabling telemetry, leaving each flag at its built-in default.
The release includes changes for teams operating the Claude apps gateway: /usage can show spend for today, this week and this month to developers without a spend limit; administrators can configure which pooled group budget pays first; and JSON gateway logs can be enabled with CLAUDE_GATEWAY_LOG_FORMAT=json. The notes also describe fixes across MCP connections, resumed tool calls, diffs, login refreshes, session handling and agent views. The page lists release v2.1.319 and its change log, but these notes do not provide a price or a separate access requirement.
What the archive adds
New Runtime’s earlier coverage on 2026-10-27 described Claude Code 2.1.311’s faster Pro/Max startup and compaction beyond 200K tokens (earlier note). On 2026-10-28, an earlier note covered v2.1.312’s gateway OIDC claim and protection for large CLAUDE.md files (earlier note). Version 2.1.319 is a new release with distinct MCP-task and auto-mode changes, not a restatement.
Trend signal
This update supports the pattern that production agents depend on runtime boundaries for safe execution: changing whether an uncheckable action prompts or is denied is a meaningful control at execution time. The fit is partial, since the release note does not describe a broader security architecture. Agent security moves to runtime boundaries.
Why now
The release notes say the Claude apps gateway now warns at startup about policies relying on coworkTabEnabled: false or chatSessionRetentionDays, ahead of a Claude Desktop change dated November 10. That gives gateway operators a specific reason to check their policies before that date.
Why it matters
Backgrounding MCP tasks changes the rhythm of tool orchestration: Claude can continue working while a long-running task proceeds, then receive its result without holding the tool call open. This can make workflows more responsive, but teams should verify that their MCP server’s task semantics and downstream steps tolerate asynchronous completion. The environment variable provides an explicit way to retain synchronous behavior where that contract matters.
The auto-mode change matters for teams routing Claude Code through gateways or disabling telemetry and feature flags. Actions the server cannot inspect no longer receive the built-in classifier as a fallback, so the resulting prompt-or-deny behavior may affect unattended workflows. Operators should review their settings and test denied or confirmation-required actions before rollout. The release also adds controls for permission prompts and gateway spend visibility, but it does not quantify reliability, latency gains or cost savings.
Limits and uncertainty
Anthropic’s release notes describe intended behavior and list fixes, but they do not report independent measurements of MCP task latency, completion reliability or the frequency of auto-mode actions that cannot be checked. The notes do not quantify how often the new prompt-or-deny path will affect gateway users. The impact will depend on each team’s MCP servers, auto-mode settings, managed configuration and permission policies; teams should validate those interactions in their own environments.
Skill to build now
Test one long-running MCP task in a non-production Claude Code session: confirm that the task moves to the background, that Claude receives its result on completion, and that downstream steps use the completed result rather than assuming synchronous execution. Then repeat with CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND set to confirm whether your workflow depends on waiting inline.
Reader takeaway
Claude Code 2.1.319 makes MCP tasks asynchronous by default and removes a safety fallback in some auto-mode configurations. Check the task behavior and gateway settings that your workflows rely on before adopting the release broadly.