CONFIRMED · OFFICIAL

Claude Code 2.1.318 reinforces credentials and MCP
Claude Code 2.1.318 surfaces login expiration, adjusts model forks and fixes issues in MCP, settings and sessions.
Marcelo Santiago
Criador do bridgecode
Get AI articles in your inbox
Technical curation on models, agents and coding tools. No spam, unsubscribe anytime.
On November 3, 2026, at 21:01 Brasília time, Anthropic published Claude Code v2.1.318 to the official repository.
The update improves model fork controls, credential visibility, sessions, settings and MCP servers. The full changelog is on GitHub.
The announcement does not introduce a new Claude model or a benchmark result. The focus is to make the terminal agent operation more predictable, especially in automations, shared environments and integrations that depend on credentials.
In this article you will see:
- How Claude Code now signals login expiration
- What changes in
effortfor model forks and inassume_role - Which MCP, configuration and session failures were fixed
- How to validate the version without putting a critical flow at risk
Claude Code 2.1.318 shows when the credential will expire
The version now reports the expiration date of a saved claude.ai login in claude auth status. The same information appears in the JSON results of claude -p with the field credential_expires_at.
For the text output of claude -p, the release adds a warning in stderr ten days before the saved login expires.
stderr is the separate channel used by programs for diagnostic messages. The main output can continue to be consumed by an automation.
The warning does not extend or renew the login. The note does not define how each CI system should store credentials, so this process continues to follow team policies.
The previous version fixed a case where claude auth status could end before saving a slow token update.
The analysis of Claude Code 2.1.317 explains this change. v2.1.318 adds the expiration date and an early warning in non-interactive mode.
Model forks receive more specific controls
The changelog adds the option effort to $.model.fork, inside a mod. A helper job can use low effort even when the main session is using another level.
The result reports which effort was sent. This allows separating a secondary task from a session that requires more careful analysis.
A fork is a separate execution started from an existing session or flow. The change allows configuring a secondary task without changing the entire session.
A quick check and a careful analysis can have distinct needs. The changelog does not state that the option improves quality or speed by itself.
The release does not list the accepted values for effort nor does it present a complete configuration example. Do not infer syntax beyond the option indicated.
Before editing a mod, check the documentation for the installed version and validate the result in a non-critical task.
Another change adds assume_role to the mantle upstreams of the Claude apps gateway. It allows using a function from another AWS account and attributing costs per developer.
It is an adjustment for administrations that already operate this gateway, not a general configuration for every local Claude Code installation.
The 2.1.316 release, about subagent effort and shell processes, addressed a related control, but with a different scope.
Image: GitHub, card of the official page of the Claude Code repository.
Claude Code 2.1.318 fixes failures in MCP and configuration files
Several fixes prevent commands from reporting success without saving changes. If a configuration cannot be updated in a shared network directory, commands no longer hide the failure.
This includes /logout, claude auth logout and claude mcp remove, according to the official changelog.
claude mcp add and removal now respect file locking. If another process is writing, the operation waits or ends with an error. The note warns that a concurrent write can still overwrite a saved configuration.
The release fixes MCP servers that shut down during version checks. They are no longer tested on every startup and can use the legacy handshake for seven days.
The rule also covers servers that shut down in less than three seconds. MCP is a protocol for connecting agents to tools and external sources.
There are fixes for automations that resume sessions. A run with --continue could lose part of the conversation after exhausting retries on a malformed tool call.
Another adjustment prevents re-run workflows from reopening agents that are already completed.
On macOS, sessions with many Bash commands could stall while memory usage grew. The changelog includes a fix for this scenario and adjustments to permissions, plugins, headless calls and remote sessions.
These are point fixes described in the release, not a guarantee that all slowness or external failures will disappear.
How to evaluate Claude Code 2.1.318 without affecting production
Identify which parts your team uses: saved login, claude -p, mods with forks, gateway, network configuration directories or your own MCP servers. The update does not affect all of these flows.
In a test environment, check claude auth status and see if the expiration date appears.
If an automation consumes the JSON from claude -p, test how it handles credential_expires_at, including when the field is missing in earlier versions.
Monitor stderr and the main output separately so diagnostics are not confused with the model response. Do not put tokens in shared logs and review log retention.
If you maintain mods, test the fork on a reversible task and check the effort reported in the result.
Since the release does not document possible values, do not copy a third party configuration without validating the official documentation.
Test adding and removing an MCP server in a disposable copy of the configuration. Check file concurrency and server shutdown on startup, without first using the configuration shared by the team.
The release does not describe a new install command. Use the procedure approved by the team, consult the official notes and do a gradual rollout, recording environment and result.
Frequently asked questions about Claude Code 2.1.318
What changes in Claude Code 2.1.318?
The release surfaces credential expiration, adds controls for model forks and fixes failures in settings, MCP and session resume.
How does Claude Code warn that login is about to expire?
`claude auth status` and the JSON output from `claude -p` now show expiration information. The text output from `claude -p` gets a warning ten days before.
What is credential_expires_at in Claude Code?
It is the field in the JSON output from `claude -p` that shows the expiration date of a saved claude.ai login.
What does the effort option do in $.model.fork?
It lets you set an effort level for a model fork, different from the one used by the main session. The release does not list the accepted values.
Does Claude Code 2.1.318 launch a new model?
No. The changelog describes changes to the CLI, mods, gateway, MCP and sessions, without announcing a new Claude model.
Claude Code 2.1.318 prioritizes operational predictability
v2.1.318 does not change the model that responds to requests. It adds expiration signals, more granular controls for forks and fixes in configuration, MCP and task resumption.
For those who maintain agents in automation, the central gain is being able to better observe credentials and failures before they affect the flow. It is still necessary to test permissions, protect files and validate integrations in the real environment.
It is not just an agent in the terminal. On bridgecode you orchestrate AI agents in a squad to split planning, implementation and review.
Source: Anthropic
Related articles

Karpathy builds a video platform over a weekend with Claude
Andrej Karpathy built HackerTube in a weekend with Claude. What it says about development speed with coding agents.

Rich Sutton reinforces that compute alone does not create intelligence
Turing Award winner Rich Sutton reinforces that raw compute is not enough for intelligence and calls for new methods, continuous learning and world models.

AI auditors detect misbehavior but generate false alarms
AI models catch a good portion of other agents' misbehavior, but they generate a lot of false alarms. See the numbers from Andy Hall's study.
