Claude Code 2.1.319: Background MCP and Permissions

In this article you will see:

• MCP tasks in the background in the terminal and how to control this behavior.
• The option permissions.denyInsteadOfPrompting to deny instead of asking.
• The command /claude-api latency-optimize to diagnose latency.
• New environment variables for feature flags and telemetry.
• Relevant fixes for users of LLM gateways.

MCP tasks in the background in the terminal

When an MCP server executes a tool call as a task, Claude Code now moves that task to the background immediately and notifies Claude when it finishes. This prevents a long call from blocking the terminal.

If you prefer Claude Code to wait for the result inside the tool call, use the variable CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND. With it active, behavior returns to synchronous.

permissions.denyInsteadOfPrompting

The new option permissions.denyInsteadOfPrompting changes permission behavior. Instead of pausing and asking the user, the tool call is denied and the turn continues. This applies to terminal sessions, subagents and teammates.

It is useful in automated flows, where a permission pause would stall the agent. Denial follows the configured permission rules, so it does not allow anything that was already blocked.

/claude-api latency-optimize

The new command /claude-api latency-optimize analyzes where your Claude API code makes the user wait and tests faster configurations against your evaluation. It is a practical diagnosis for those building on the API who want to reduce latency without losing quality.

New environment variables

Three new variables appear in this release. CLAUDE_CODE_DISABLE_FEATURE_FLAGS stops fetching feature flags without turning off telemetry, using the built-in default for each flag. CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND controls the synchronous behavior of MCP tasks. CLAUDE_CODE_AUTO_MODE_SERVER enables server-side checking even with CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS active.

For Claude apps gateways, there is also CLAUDE_GATEWAY_LOG_FORMAT=json, which transforms operational log lines into JSON with severity, time and message fields.

Fixes for LLM gateway users

Several adjustments affect users running Claude Code behind a gateway. Auto mode now warns when server-side verification does not reach the session. /status no longer shows "Auto mode server: Enabled" in cases where verification is not active. The model no longer tries to send feedback on third-party gateways, where it cannot be delivered.

A case was also fixed where Bash deny and ask rules did not capture commands written with words in quotes.

How to test today

The simplest flow is: update to 2.1.319, test a long MCP call to see the background behavior, and enable permissions.denyInsteadOfPrompting in a test project to measure the impact on automated flows. For those building on the API, run /claude-api latency-optimize in a real project.

Conclusion

2.1.319 focuses on agentic flows: MCP tasks in the background, permissions that do not block and latency diagnosis. For those using Claude Code in automated pipelines, these changes reduce stalls.

You define the goal and the agents split the tasks: one plans, another executes, another reviews. See how orchestration works at bridgecode.

Fonte: Anthropic

Image and references

See also claude code 2 1 318 reinforces mcp credentials, mcp server git from mcp for ai clis, claude code 2 1 318 reinforces mcp credentials and mcp server git from mcp for ai clis.