Claude Code September updates add managed controls, Sonnet 5.5, and Mods

Claude Code’s September 2026 releases add several controls for teams running coding agents in managed or unattended environments. Versions 2.1.257 through 2.1.260 bring Fable 5.1 support, organization-managed MCP servers, a headless permission mode, and a new fullscreen diff panel.

Centrally managed MCP servers

Version 2.1.259 adds the managedMcpServers managed setting. Organizations can provide HTTP or SSE MCP servers to every user using the same entry shape as .mcp.json; entries that would run a local command are skipped.

The release also changes how allowedMcpServers is applied: it governs servers users add, while a literal managed MCP entry is loaded unless a deniedMcpServers rule keeps it off. Administrators should review both settings during migration so a policy does not unintentionally expose or suppress a server.

Headless runs can deny prompts instead of hanging

The new --permission-prompts none option is designed for unattended headless hosts. Anything that would require a permission prompt is denied automatically, while the active permission mode continues to decide actions that do not need a prompt.

That behavior is safer for automation than waiting forever for a person who is not present, but it can also make a job finish without completing its intended work. Treat a denied prompt as a structured failure in orchestration and surface it in job logs.

Fable 5.1 and review visibility

Claude Code 2.1.257 adds Claude Fable 5.1 as the default Fable model, with a 1M context window, adaptive thinking, and the documented $10/$50 per-million-token input/output rates plus cheaper cache reads. Version 2.1.260 adds a fullscreen diff panel that opens beside the conversation and shows uncommitted changes as Claude edits; toggle it with /diff.

The same release adds text-form /advisor commands for desktop, Remote Control, and headless sessions, plus more diagnostic detail for prompt-cache misses. These features make it easier to inspect and steer long runs, but teams should still review the actual diff and tool permissions before merging.

September 4–10 follow-up: gateway, effort, and output controls

The following Claude Code changelog entries extend the earlier 2.1.257–2.1.260 coverage. The September 10 update adds pricing information to the Claude apps gateway so signed-in clients can match managed rates in /cost and telemetry, warns when a gateway has no allowed CIDR ranges, and adds gatewayInternalNetworks for organizations that want gateway login from their own public IPv4 block.

The same entry adds claude self-hosted-runner --remove-session-state, which can delete per-session directories when a runner session ends, and JSON output for plugin install, uninstall, update, enable, and disable commands. On September 9, Claude Code added maxEffortLevel to cap model effort across providers and --system-prompt-snapshot off for refreshing the system prompt on every request.

These are operational controls rather than a new model launch. Administrators should test gateway network policy, runner cleanup, plugin automation, and effort caps in a non-production environment before changing managed settings across a fleet.

Claude Managed Agents session viewer

In a September 11 announcement on the official ClaudeDevs account, Anthropic said Claude Managed Agents added a session viewer to the ant CLI. The post says ant beta:sessions connect attaches a terminal to a running session and that --web opens a web UI served from localhost. This section is limited to what the official announcement states; consult current Managed Agents documentation before relying on the beta command.

September 18–24 follow-up

Claude Code 2.1.277, released September 18, adds AGENTS.md support: when a project has no CLAUDE.md, Claude Code reads AGENTS.md instead. The setting can be changed under Project instructions in /config; the changelog notes that this is not yet available on Bedrock, Vertex, or Foundry.

Version 2.1.278 changes auto mode for Claude API and Enterprise users and for Bedrock, Vertex, Foundry, and gateways to use the server-side classifier by default. The documented opt-out is CLAUDE_CODE_AUTO_MODE_SERVER=0 on the listed hosted providers, and the release warns when a billed fallback is used.

Version 2.1.282, dated September 24, adds maxProseWidth for wide terminals while leaving tables and code blocks at full width. It also adds notices for ignored or disabled telemetry variables, a managed allowClaudeInChromeWithManagedMcp setting, and clearer Cloud sessions GitHub App status in connector settings. The release adds attribution: false in settings.json to hide commit and pull-request attribution and adds MCP URL-mode elicitation for supported protocol connections.

These updates affect project-instruction discovery, provider billing, managed settings, and repository attribution. Test the precedence between CLAUDE.md, AGENTS.md, and managed settings in a disposable repository before changing shared templates, and review provider-specific auto-mode billing before enabling it across a fleet.

Reliability and migration notes

The releases fix concurrent-session corruption in ~/.claude.json, resume failures involving attachments or worktrees, managed-settings enforcement gaps, and several MCP, permission, and remote-session edge cases. The 2.1.260 changelog also reverses a 2.1.259 change that applied Read() deny rules to Bash arguments too broadly, so organizations with those rules should test their automation after upgrading.

Because these are fast-moving CLI releases, pin and test the version used in CI or self-hosted runners. Validate managed MCP policy, unattended prompt behavior, model selection, prompt caching, and diff review before rolling the update across a fleet.

September 17–18 follow-up

Claude Code 2.1.275 adds account confirmation for gateway sign-in, a send-now shortcut, syncing for skills and plugins enabled on a claude.ai account, and marketplace-aware plugin installation. Version 2.1.276 fixes a regression that caused every request to fail when ANTHROPIC_BASE_URL pointed to a proxy or gateway.

These are operational and workflow improvements rather than a new model release. Teams using a gateway should test the 2.1.276 fix and review whether account, skill, and plugin synchronization matches their organization’s policy.

The official ClaudeDevs account also said on September 16 that Claude Design, Claude Slides, and Claude Docs work inside Claude Code. The announcement says users can point those experiences at files and RFCs in a repository to create a design review deck or UI mockup, then edit it or continue the conversation. This paragraph is limited to the official announcement; availability and plan coverage should be confirmed in the product before relying on it.

September 25–October 1 follow-up: model controls, plugins, and Mods

Claude Code 2.1.283 adds prompt IDs to gateway hint headers and managed availableModelsMatch and deniedModels settings, so an organization can require an exact listed model version or block a named model. The release also adds /doctor prompt-audit to check project instructions, skills, agents, and commands for prompting patterns written for older models.

Version 2.1.284 adds Claude Sonnet 5.5 as the default Sonnet model on the Anthropic API. The changelog lists a 1M-token context window and input/output pricing of $2/$10 per million tokens, with $0.20 per million cache reads. It also adds an auto-mode option to allow a read outside the working directories once while still asking again for later reads.

Version 2.1.285 adds CLAUDE_CODE_DISABLE_WEB_FETCH to turn off WebFetch, claude --desktop to open the desktop app on the current directory or a selected session, and claude plugin configure <plugin> to inspect or save plugin options. 2.1.286 adds a count to stacked permission prompts and mouse support for the extra rows in fullscreen lists.

Version 2.1.287 introduces Claude Mods, which let plugins modify deeper behavior, and a built-in You should know mod that uses a side agent to flag things the user or Claude might miss. The release says to enable it with /plugin enable cc-plugin-you-should-know@builtin; it is for first-party sessions with telemetry on. The official Mods guide says Mods require v2.1.287 or later and are on by default; it documents support in the CLI and the Code tab of the Claude Desktop app.

These additions touch model access policy, file-reading permissions, plugin behavior, and telemetry. Test managed model matching, the WebFetch setting, and any installed Mods in a non-production environment before distributing shared configuration.

Sources

• Claude Code Docs — “Changelog”: https://code.claude.com/docs/en/changelog
• Anthropic — Claude Code official repository changelog: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md
• Anthropic — Claude Code releases: https://github.com/anthropics/claude-code/releases
• Anthropic — Claude Code release v2.1.287: https://github.com/anthropics/claude-code/releases/tag/v2.1.287
• Claude Code Docs — “Mods overview”: https://code.claude.com/docs/en/plugins/mods/overview
• ClaudeDevs on X — official developer updates, including the September 11 Managed Agents session-viewer announcement: https://x.com/ClaudeDevs
• ClaudeDevs on X — Claude Design, Claude Slides, and Claude Docs inside Claude Code (September 16, 2026): https://x.com/ClaudeDevs/status/2100270861555228770