Claude Code 2.1.319: MCP Tasks Go Background, Release Notes
Anthropic released Claude Code 2.1.319 on November 6, 2026, according to the official changelog. The release targets the two points where unattended coding agents stall. Long MCP tool calls no longer hold the session, because tasks move to the background by default, and a new setting can turn permission prompts into denials so a turn keeps moving instead of waiting for a person who is not there.
The other half of the release pulls in the opposite direction. Auto mode now stops when its safety checks cannot reach the work, with no classifier fallback behind an LLM gateway. Read together, the defaults say a session should never sit idle waiting, and it should never proceed unverified either. For teams running Claude Code as automation rather than as an interactive assistant, those two defaults matter more than any single bug fix in the long changelog tail.
MCP Tasks in the Background
When an MCP server executes a tool call as a task, Claude Code now hands it to the background immediately and notifies Claude when the result is ready. MCP Tasks themselves arrived earlier, as a public beta covered in issue #3525 on October 23, with a structure where a server takes on long work and returns the result later. The beta attached extra conditions to the background handoff. Version 2.1.319 removes that friction and makes the background path the default behavior.
Workflows that genuinely need the old blocking behavior keep an escape hatch. Setting the environment variable CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND makes the task finish inside the tool call, the way it did before. That split is the practical shape of the change: parallel, fire-and-continue work becomes the norm, while sequential steps that depend on a result can still opt back into waiting with one variable.
Permission Prompts Turned Into Denials
The new permissions.denyInsteadOfPrompting setting denies tool calls that would otherwise wait at a permission prompt, and the turn continues past them. It applies across terminal sessions, subagents and teammates, so an unattended fleet does not stall one worker at a time. The direction matches the previous release line: version 2.1.317, covered in issue #3731, added Agent SDK hooks that block on failure, and 2.1.319 now covers the prompt itself.
The trade-off is visibility. The issue notes that with the setting on, a session makes it hard to tell why a particular call was denied. In an interactive session that opacity is simply frustrating, because the person at the keyboard could have approved the call in seconds. The setting is therefore best read as an unattended-run control, not as a general replacement for prompts, and teams enabling it need their own logging to reconstruct what the agent skipped.
Auto Mode Changes Behind LLM Gateways
Auto mode changes its default in environments where telemetry or feature flags are turned off. Behind an LLM gateway, calls the server cannot verify are now prompted or denied, and there is no fallback that substitutes the built-in classifier. Keeping the previous behavior takes CLAUDE_CODE_AUTO_MODE_SERVER=0. Bedrock, Vertex and Foundry are explicitly not affected by the change.
The reverse switch also exists. Setting CLAUDE_CODE_AUTO_MODE_SERVER=1 turns on server-side checks behind an LLM gateway, and it applies even when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set. A related addition, CLAUDE_CODE_DISABLE_FEATURE_FLAGS, stops feature flag fetching without turning off telemetry, leaving every flag on its built-in default value. For gateway operators the message is direct: check the CLAUDE_CODE_AUTO_MODE_SERVER configuration before upgrading, because existing automation can change behavior the moment the new default lands.
claude -p run that is denied in auto mode because no safety check was available now ends with exit code 1. It previously ended with exit code 0, so wrappers that treated a clean exit as success will now, correctly, see these runs as failures.Usage, Gateway Budgets and Session Tools
The remaining additions cluster around operating Claude Code at team scale. The /usage command now shows spending for today, this week and this month, aimed at developers behind a Claude apps gateway who have no spending limit, and it requires the gateway to run this version or later. Gateways also gain pooled budget rules, admin.pooled_conflict_rule and admin.pooled_priority_order, which decide which group budget pays first when a developer belongs to several groups. A login_notice setting puts an operator message on the login screen, and CLAUDE_GATEWAY_LOG_FORMAT=json writes operation logs as JSON.
Two smaller tools round out the release. The new /claude-api latency-optimize command finds the points where a user waits in Claude API code and tests faster settings against an evaluation set. In /tasks, an Earlier in this session line lets you reopen a finished subagent and send it another message, including after a session restart. Cloud sessions also get a clearer warning in the /loop check: the timer stops together with the container, and the check suggests a persistent loop where that matters.
Security Fixes in the Release
The changelog lists several security fixes that sit close to the permission theme. A sandbox escape during worktree creation is fixed. A bypass that hid Bash permission checks behind an environment prefix is blocked. Cases where sed skipped permission checks in accept-edits and auto mode are also fixed. None of these change a workflow on their own, but together with the prompt and gateway changes they show where this release cycle is spending its effort: closing the gaps between what the permission system intends and what a determined command line can route around.
Limits and Unconfirmed Details
One protocol question stays open. The beta documentation for MCP Tasks required protocol revision 2026-07-28 on supporting MCP servers, and the 2.1.319 changelog does not say whether that requirement changed. Servers built against the beta should keep working on that assumption until Anthropic documents otherwise. The changelog is also long, and the issue assessment is blunt about its shape: most entries are individual bug fixes, and the changes users will actually feel are the background tasks, the denial setting and the auto mode defaults covered above.
Related Reading
For the earlier step in this same release line, see the report on Claude Code background subagents and the 99 percent status cost cut. For another take on moving agent execution away from the local machine, see Antigravity Remote Agents and cloud coding sessions.
FAQ
What changed for MCP tasks in Claude Code 2.1.319?
When an MCP server runs a tool call as a task, Claude Code now moves it to the background right away and notifies Claude when it finishes. Setting CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND restores waiting inside the tool call until the task ends.
What does denyInsteadOfPrompting do?
The new permissions.denyInsteadOfPrompting setting denies tool calls that would otherwise wait at a permission prompt, and the turn continues. It applies to terminal sessions, subagents and teammates.
How did auto mode change behind LLM gateways?
In environments where telemetry or feature flags are off, auto mode now prompts or denies calls the server cannot verify, with no fallback to the built-in classifier. CLAUDE_CODE_AUTO_MODE_SERVER=0 keeps the previous behavior, and Bedrock, Vertex and Foundry are not affected.
What is still unconfirmed in this release?
The beta documentation for MCP Tasks required protocol revision 2026-07-28, and the 2.1.319 changelog does not say whether that requirement changed. The changelog also leaves open how clearly a session shows why a call was denied under denyInsteadOfPrompting.