Claude Code Moved MCP Tasks to the Background. Read the New Deny Switch First.

Anthropic shipped Claude Code 2.1.319 on 7 November. It is a maintenance release with three behaviour changes, and the one that will get discussed is not the one that matters most.
The headline: MCP tool calls that run as tasks now execute in the background in the terminal, and Claude is told the result when they finish. Before this, a long MCP call blocked the session. Anyone who has watched a browser automation tool sit on a page for ninety seconds while the prompt spinner turns gets the appeal. A variable makes the call wait again if you prefer the old behaviour; the release notes mention it without putting it front and centre.
Background execution is the kind of feature that is easy to like and hard to reason about. A session that used to do one thing at a time now has work in flight. What happens when a background task finishes during a permission prompt? What happens when two background tasks touch the same file? The release does not answer those, and the answers will show up as bug reports over the next month.
The setting that changes who is responsible
permissions.denyInsteadOfPrompting does something narrower and more consequential. When set, a tool call that would otherwise open a permission prompt is refused instead.
Read that twice. The default behaviour of an agent permission system is to ask a human. That works when a human is at the keyboard, which is the assumption every coding agent has shipped with. It stops working the moment you run the agent from a scheduler, a CI job, or an overnight batch: there is nobody to answer, and depending on the implementation you either block forever or the prompt times out into something that is not obviously a denial.
denyInsteadOfPrompting turns the unanswered question into a clean no. The call fails, the agent sees a denial it can reason about, and the run continues or stops on a decision rather than a timeout.
That is the difference between an agent that can be left alone and an agent that cannot. It is also a setting you should think about before you turn on, because it inverts a safety default: the system stops asking and starts refusing. For a read-only or well-scoped workload, fine. For an agent with write access to a production system, refusing is the right instinct and the refusal has to be observable, which is where the JSON log output in the same release becomes useful.
Auto mode got stricter about what it cannot see
The third change is the one with the longest tail. Auto mode, per the release notes, now behaves differently when telemetry or feature flags are off, including behind an LLM gateway: calls it cannot verify are asked about or denied outright, with no fallback to the built-in classifier.
The logic is sound. A classifier that decides whether a command is safe is a safety mechanism, and if the environment prevents it from working, pretending it worked is worse than admitting it did not. Teams running Claude Code behind a corporate gateway or with telemetry disabled should expect more prompts, or more denials if they have set the new flag. That is a correct trade and an annoying one, and it will produce a small wave of "why is it asking now" issues from exactly the organisations that value the privacy settings.
The two additions that are just useful
/claude-api latency-optimize is a command, presumably for tuning API call patterns toward lower latency. The notes do not explain the mechanism, so treat it as a knob to try rather than a fix.
/usage now reports spend for the Claude apps gateway, and the release adds pooled-budget conflict rules. Both are accounting features, and accounting features are how agent platforms become acceptable to finance departments. A team that cannot attribute spend per project does not get to scale a coding agent, regardless of how good it is.
Also in the release: JSON log output, and fixes across permissions, MCP, cloud sessions and the VS Code extension. And a behaviour change worth flagging for anyone who scripted around it: self-paced /loop now ends when a turn finishes rather than rescheduling itself, unless CLAUDE_CODE_LOOP_KEEPALIVE=1 is set. If you had a loop running as a heartbeat in the background, it stopped.
Why a point release deserves a post
Because the three changes compose into a policy. Background execution means more work in flight. denyInsteadOfPrompting means fewer chances for a human to intervene. The auto-mode change means the classifier cannot be counted on in restricted environments. Together they describe an agent that is expected to run without supervision, in places where the network is filtered and the logs are shipped somewhere else.
That is a different product from the one that shipped a year ago, and the shift is happening through release notes rather than announcements. Which is normal. Infrastructure products mature by accumulating defaults, and the defaults are where the real opinions live.
There is a practical point about background tasks that the notes leave implicit. A tool call that used to block is now a thing you have to wait for, which means the session needs a way to represent "not finished" that an operator can read. In a terminal that is fine. In a scripted pipeline, it is the difference between a step that completes and a step that returns an empty result because the tool was still running. Anyone wrapping Claude Code in a shell script should assume the timing changed and test for it.
The same applies to denyInsteadOfPrompting in reverse. A denial is a tool result, not a crash, and an agent that receives a denial may try a different route. That is usually what you want and occasionally a way to route around a boundary you set on purpose. If the boundary matters, it belongs in the permission rules, not in an expectation that the agent will stop at the first no.
For anyone running Claude Code unattended, the checklist from this release is short. Decide whether background MCP tasks are acceptable for your workload, or pin the wait variable. Decide whether a denial or a prompt is the right failure mode for each job, and set denyInsteadOfPrompting per environment rather than globally. Check that your gateway setup still lets auto mode verify what it needs, and if it does not, budget for more approvals. And if anything depended on a self-paced loop rescheduling, set CLAUDE_CODE_LOOP_KEEPALIVE=1 before the next run.
None of that is exciting. All of it is the difference between an agent that fails loudly at 3am and one that fails silently at 3am, and only one of those is recoverable by morning.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.