Claude docs changes for September 10th, 2026 [diff]

Executive Summary

• The Admin API (organization members, RBAC groups/roles, MCP tunnels, spend limits, cost/usage reports, and Enterprise analytics) is now also exposed as a typed beta SDK surface — client.beta.organization in Python, TypeScript, C#, Go, Java, PHP, and Ruby, and ant beta:organization in the CLI — documented under a new api/beta/organization/ reference tree alongside the existing REST-only Admin API docs, and every guide that used to link to api/admin/* now points at the new location.
• Claude Managed Agents permission policies gain a third option, auto: the server evaluates each agent or MCP tool call and runs it, denies it, or pauses it for your approval, reporting the outcome on agent.tool_use/agent.mcp_tool_use events via new evaluated_permission and evaluation fields.
• New ant beta:sessions connect CLI command attaches your terminal (or, with --web, the Claude Console's session viewer) to a live Managed Agents session, so you can follow its transcript, send messages, interrupt it, and allow or deny tool calls waiting for approval.
• Claude Code artifact publishing now validates the source file decodes as UTF-8 (or UTF-16 with a BOM) and refuses ones that don't, or that contain a stray U+FFFD, naming the exact line and column to fix; image-resize failures also now name specific causes such as a CMYK JPEG or animated WebP instead of a generic decode error.
• Effort caps (maxEffortLevel and organization-wide caps) now also constrain skill/subagent frontmatter effort and gate whether ultracode is available at all; allowedChannelPlugins gains a "plugin@marketplace" string shorthand, and managed allowedHttpHookUrls/httpHookAllowedEnvVars now fail open to lower-scoped settings instead of blocking everything when invalid.

Claude Code changes

Changed documents

agent-sdk/typescript [Source]

• Clarified that effortLevel: "ultracode" requests xhigh effort with ultracode on, rather than guaranteeing it runs at xhigh — it can still be capped. [line 586] [Source]

artifacts [Source]

• The published source file must now decode as UTF-8, or as little-endian UTF-16 by its byte-order mark; a file that doesn't decode, or that contains the replacement character U+FFFD, is refused with the line and column to fix instead of being published. [line 280] [Source]

cli-reference [Source]

• --effort ultracode now requests xhigh effort with ultracode on, rather than guaranteeing it starts there — an effort cap or unsupported model can still lower it. [line 81] [Source]

errors [Source]

• Added a new "The source file is not valid UTF-8 text" section (v2.1.267+): publishing an artifact from a file that doesn't decode as UTF-8/UTF-16, or that already contains U+FFFD, is refused before upload, naming the first bad position; before this version the server refused the publish instead. [lines 212-213, 3092-3112] [Source]
• Added a new "Working directory no longer exists" background-session error (v2.1.257+), shown when dispatching or restarting a session whose working directory was deleted or moved; it previously appeared to start and then failed silently in agent view. [line 232, 3423-3435] [Source]
• Image-resize failure messages now name a specific cause when known — a CMYK JPEG, an animated WebP, or a possibly damaged file — and say what format to re-save as, instead of a generic "processing is unavailable" message. [lines 1690-1701] [Source]

ide-integrations [Source]

• The model-name button in the prompt box footer now shows the selected effort level (v2.1.257+). [line 95] [Source]
• You can now paste an image from your clipboard directly into the prompt box, in addition to drag-and-drop. [line 124] [Source]
• Clicking a session in history that's already open in another tab of the current window now switches to that tab instead of opening a duplicate. [line 130] [Source]

managed-settings [Source]

• allowedHttpHookUrls and httpHookAllowedEnvVars now merge across settings files when the managed value is invalid: entries from user, project, or local settings still apply while the managed list falls back to empty, rather than blocking everything (v2.1.267+; earlier versions drop the whole key). [line 329] [Source]

model-config [Source]

• Spelled out exactly when ultracode is unavailable — workflows turned off, the model doesn't support xhigh effort, or an effort cap below xhigh applies — and that --effort ultracode now starts the session at the highest allowed effort with ultracode off in those cases, rather than only covering the workflows-off case. [lines 558-566] [Source]
• Noted that a maxEffortLevel or organization effort cap still limits the level a skill's or subagent's frontmatter effort can run at. [line 599] [Source]

monitoring-usage [Source]

• Clarified that the user.email telemetry attribute is sent only to the OTel endpoint you configure, never to Anthropic, matching the existing guarantee for other OTel fields. [line 1331] [Source]

remote-control [Source]

• Documented that when a claude remote-control server's registration credential expires, it re-registers with the Anthropic API automatically and keeps serving its sessions. [line 218] [Source]

settings-reference [Source]

• maxEffortLevel now documents that a cap below xhigh also makes ultracode unavailable on the models it applies to. [line 958] [Source]
• allowedChannelPlugins entries can now be written as a "plugin@marketplace" string (e.g. "telegram@claude-plugins-official") instead of an object; the string form requires v2.1.267+, and earlier versions reject the whole value if one is used. [line 4033] [Source]

vs-code [Source]

• Received the same model-effort-badge, clipboard-image-paste, and open-tab-switching updates as ide-integrations, since this page mirrors that content. [lines 95, 124, 130]

workflows [Source]

• The /effort menu now links to the fuller "when ultracode is available" explanation (workflows off, unsupported model, or an effort cap) instead of only mentioning xhigh-effort model support. [line 148] [Source]

API changes

New Documents

api/beta/organization/analytics [Source]

The Claude Enterprise Analytics API — activity summaries plus per-area usage for artifacts, chat projects, connectors, cost, plugins, skills, usage, and users — is now documented as a typed beta SDK endpoint group (client.beta.organization.analytics.* in Python/TypeScript/C#/Go/Java/PHP/Ruby, ant beta:organization), functionally identical to the existing Admin API analytics reference and requiring no beta header beyond the standard anthropic-version.

api/beta/organization/cost_report [Source]

Beta-typed mirror of the Admin API's service-level cost report endpoint, grouping USD cost by workspace or description at daily granularity.

api/beta/organization/mcp_tunnels [Source]

Beta-typed mirror of the Admin API's organization-scoped MCP Tunnels endpoints (list, retrieve, archive, reveal/rotate token, and certificate management). The doc marks this org-scoped surface deprecated: new integrations should use the dedicated /v1/tunnels API (anthropic-beta: mcp-tunnels-2026-06-22) instead.

api/beta/organization/rbac_groups [Source]

Beta-typed mirror of the Admin API's RBAC Groups endpoints (list, get, create, update, delete, and member management) for Claude Enterprise organizations; requires the anthropic-beta: ce-user-management-2026-07-13 header.

api/beta/organization/rbac_roles [Source]

Beta-typed mirror of the Admin API's read-only custom-roles catalog (list roles, get role, list role permissions) for Claude Enterprise organizations.

api/beta/organization/spend_limits [Source]

Beta-typed mirror of the Admin API's per-user spend limit endpoints (set, get, delete, list effective limits) plus spend-limit increase requests (list, get, approve, deny).

api/beta/organization/usage_report [Source]

Beta-typed mirror of the Admin API's Messages and Claude Code usage report endpoints.

cli-sdks-libraries/cli/sessions-connect [Source]

New page for ant beta:sessions connect, which attaches your terminal to a live Managed Agents session: it loads and follows the transcript, and lets you send messages, interrupt the agent, and allow or deny tool calls waiting for approval. --web instead serves the Claude Console's session viewer locally and opens it in your browser, following every thread of a multiagent session.

Changed documents

manage-claude/admin-api [Source]

• Documents that the Python, TypeScript, C#, Go, Java, PHP, and Ruby SDKs now expose the Admin API under client.beta.organization, and the ant CLI under ant beta:organization. [line 29] [Source]

manage-claude/user-management [Source]

• The "Remove user" endpoint reference now points at the beta org API's users/remove.md, distinct from the old admin users/delete.md naming. [line 159] [Source]

managed-agents/events-and-streaming [Source]

• Tool confirmation now also pauses under the new auto policy when the server reaches no determination, not just under always_ask; agent.tool_use/agent.mcp_tool_use events carry a new evaluated_permission field and, usually, an evaluation object recording which policy produced the outcome, shown in a worked example. [lines 2143-2150] [Source]
• Noted that ant beta:sessions connect can open the same Console session viewer from the CLI or follow a session in your terminal. [line 2774] [Source]

managed-agents/migration [Source]

• Updated the SDK-migration comparison tables to mention auto as a third permission_policy option, and to note that under auto a call the server evaluates as safe runs without ever reaching your client (unlike always_ask, which always does). [line 620, 1297]

managed-agents/permission-policies [Source]

• Added a new auto permission policy: the server evaluates each agent or MCP tool call and either runs it (safe), denies it (high-risk, with no client override), or pauses it for approval (no determination reached). Set it via permission_policy: {"type": "auto"} on a toolset's default_config or a per-tool configs entry; no toolset uses it by default. Documented with a full multi-language example that sets auto as the default for an agent toolset and an MCP toolset while overriding bash to always_ask. [line 19, 654-663] [Source]

managed-agents/reference [Source]

• agent.tool_use and agent.mcp_tool_use event descriptions now note they carry evaluated_permission and, usually, evaluation. [lines 34, 36]

release-notes/overview [Source]

• Added a September 10, 2026 entry for the auto permission policy and ant beta:sessions connect. [lines 15-18] [Source]

A note on the rest of this diff

Beyond the entries above, this sync touched roughly 1,500 files, almost all of it the same one-line change repeated: every guide and per-language SDK reference page (Python, TypeScript, C#, Go, Java, PHP, Ruby, and the CLI, across api/admin, api/beta, and each per-language tree) that used to link to an api/admin/* reference page now links to the equivalent api/beta/organization/* page instead. Per the "ignore small URL changes" rule, those are omitted here individually; the underlying reason for all of them is the new beta SDK surface described in the New Documents section above.