Skip to content

Claude Code

Claude Code 2.1.285 lets admins lock down API providers

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →

Claude Code 2.1.285, released on September 29, adds an allowedProviders managed setting that limits which API providers a machine may use. It also caps background Bash and PowerShell commands at 30 minutes by default, so a long job you send to the background now needs its own timeout. Both are in the release notes on GitHub, along with a long run of permission, plugin and retry fixes.

The provider list covers the Anthropic API, a custom endpoint, Bedrock, Mantle, Vertex AI, Foundry, Claude Platform on AWS and a Cloud gateway. Two fixes in the same release affect whether the policy loads in the first place. Sessions that authenticate with ANTHROPIC_AUTH_TOKEN against the Anthropic API never loaded the organization’s policy until now. And when the OS denies Claude Code read access to the managed settings file, it used to refuse to start; now it warns and starts without that file’s policies. Any other read error, or a file it can’t parse, still stops every session. So if the OS blocks Claude Code from reading the managed file, the session starts with none of its limits, provider list included.

A related change for admins on Bedrock and Vertex AI: when an admin removes access to the default model, sessions now switch to an older available model of the same tier instead of failing, and session titles and summaries fall back with it.

Before you update

The background limit comes from the timeout set with run_in_background: 30 minutes by default, two hours at most. Claude is notified when a command is stopped. A build or test run that takes longer than two hours can no longer run as a background command.

Sessions behind a custom ANTHROPIC_BASE_URL now use the 1M context window on models that have one: Opus 4.7 and later, Sonnet 5 and later, and Fable. If your gateway stops at 200K, the release notes say to run /autocompact 200k.

Headless runs change too. claude -p and Python Agent SDK sessions on third-party providers, or with telemetry off, now start in auto mode when no permission mode is configured, the same as interactive sessions. Version 2.1.284 had already started terminal and VS Code sessions in auto mode on every plan and provider when no mode was set. With this release, a script on Bedrock or Vertex that never set a permission mode gets auto mode as well. Passing --permission-mode still overrides it.

Some smaller changes may break existing setups. In cloud sessions and on self-hosted runners, the MCP server name widgets is now reserved: a server of your own with that name, or a close spelling such as widgets_, won’t load until you rename it. On Windows, project and local settings env can no longer set ALLUSERSPROFILE, SystemDrive or the CommonProgramFiles variables. Set those in user or managed settings instead. On macOS and Linux, /ultrareview now needs git 2.31 or newer to upload a local repository, and checkouts made with --separate-git-dir are refused.

Retries, WebFetch and new commands

Until now, a failing API request could be retried a full 21 times when streaming kept failing, because the non-streaming fallback got its own fresh set of retries. The fallback now shares the request’s retry budget. A new CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES variable caps how many times a timed-out fallback request is re-sent. Responses blocked by the API’s output content filter also stop being retried, which could go on for minutes, and the filter’s error now shows right away.

Setting CLAUDE_CODE_DISABLE_WEB_FETCH turns off the WebFetch tool. If the organization policy fails to load at startup, Team and Enterprise sessions, and sessions whose plan Claude Code can’t determine, hold WebFetch back until the policy loads. When a domain safety check is rate-limited, WebFetch now says so instead of reporting a network or enterprise policy block.

For plugins, claude plugin configure <plugin> lists a plugin’s options, flags the ones still unset, and saves new values read from stdin with --values-stdin. claude plugin install --config now accepts <server>.<key>=<value>, so a bundled .mcpb MCP server can get its settings at install time and start without a trip to /plugin → Configure. Separately, claude --desktop opens the Claude desktop app on the current directory, or on a session with --continue or --resume.

Where permission rules failed to apply

When the PowerShell tool’s command parser failed to start, for example on a machine out of memory, its permission check skipped deny and ask rules and cached that failure for later checks. A reply sent from claude agents to a background session waiting on a permission prompt sometimes approved the pending command. Fork subagents now run under their parent’s permission mode and can’t exit plan mode. In SDK and -p sessions, an MCP server added mid-session and then switched off used to leave its tools available. Project settings also lose some reach: they can’t widen or turn off a sandbox an admin requires, extend a strict allowlist, or reopen read paths that managed settings deny.

One fix may change how an existing rule behaves. An Artifact allow rule (“don’t ask again”) used to let the Artifact tool publish a file outside the working directories without asking. For the rule to cover a file elsewhere, add that file’s folder with --add-dir.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.