Claude Code 2.1.259 shipped on September 2. Not a huge release, but it carries a few changes that matter mostly to teams and to anyone running Claude Code without a human at the keyboard.
The biggest one is managedMcpServers. Organizations can now provide HTTP and SSE MCP servers to every user centrally, using the same entry format as .mcp.json. Entries that would launch a local command are skipped. If you maintain MCP servers for Jira, Confluence or internal databases at your company, you no longer have to hope every developer sets up their config correctly. Alongside that, allowedMcpServers changes meaning: the list now only governs servers that users add themselves. To keep a managed server off, use deniedMcpServers.
The second addition is --permission-prompts none. On hosts where nobody is watching, anything that would normally trigger a prompt is denied automatically. The active permission mode, including auto mode, keeps deciding everything else. For cron jobs, CI pipelines and similar runs, this is the missing switch that lets a job fail cleanly instead of waiting forever for an answer.
Smaller additions: GitLab users now see merge requests created or merged via glab mr shown as “MR !N” in the collapsed tool summary. claude plugin validate gets a --json flag for machine-readable reports. And /install-github-app, when run inside a GitLab repo, explains that it is GitHub-only and points to the GitLab CI/CD docs.
Among the bug fixes, one stands out. Until now, several sessions running at the same time could silently revert each other’s changes to ~/.claude.json. The result: workspace trust suddenly reset, or MCP and project state went missing. If you, like me, tend to run several Claude Code windows in parallel, you may know that vague feeling that something is off. That should be over now.
More fixes: Bash Read() deny rules now also cover files passed as option values, such as --ignore-revs-file=.env, file operands of git diff, and chains like cd DIR && cat FILE. The prompt cache used to be invalidated when the OAuth token refreshed in sessions with telemetry disabled. Auto mode sometimes ran a turn on a model it doesn’t support when a command or skill named a model: in its frontmatter. And Stop didn’t actually stop background agents in remote-control sessions.
One item is more of a tightening than a fix. If managed settings, a drop-in file, the MDM plist or the HKLM value on Windows can’t be parsed, Claude Code now refuses to start and names the broken source. Previously, the settings simply went unenforced in silence. For admins, that’s the right call, even if it will cause a few Monday-morning surprises.
In VS Code, the session list sidebar gets a new “Active” quick filter and a status menu to filter by “Needs input,” “Working” and “Completed.”
To me, 2.1.259 is a release for people who roll Claude Code out across a team or let it run unattended. If you work alone on a laptop, you’ll barely notice it. But that’s exactly where the tool is growing up right now.
Sources: Claude Code Changelog, Releasebot: Claude Code Updates