Claude docs changes for March 24th, 2026 [diff]

Executive Summary

• Computer Use on Desktop (macOS): Claude can now open apps and control your screen on macOS (Pro/Max plans). New documentation covers enabling it, per-app permission tiers, and safety boundaries.
• Cloud Scheduled Tasks: New cloud-based scheduled tasks run on Anthropic infrastructure and keep running even when your machine is off. Create them from the web, Desktop app, or CLI via /schedule.
• Dispatch Integration: Desktop now supports Dispatch — send a task from your phone's Claude app and a Desktop Code session spawns automatically.
• New Platforms & Integrations overview: New dedicated page comparing all Claude Code surfaces (CLI, Desktop, Web, IDE extensions) and remote access options.
• Agent SDK Tool Search: New dedicated guide for MCP tool search, which is now on by default and supports up to 10,000 tools.

Claude Code changes

New Documents

Platforms and integrations [Source]

New overview page comparing all Claude Code surfaces (CLI, Desktop, VS Code, JetBrains, Web) and integrations (Chrome, GitHub Actions, GitLab, Code Review, Slack). Includes a table for remote access options (Dispatch, Remote Control, Channels, Slack, Scheduled tasks) with trigger, runtime, setup, and best-use-case columns.

Schedule tasks on the web [Source]

New documentation for cloud scheduled tasks that run on Anthropic-managed infrastructure. Covers creating tasks from the web, Desktop app, or CLI (/schedule), selecting repositories and environments, configuring MCP connectors, frequency options (hourly, daily, weekdays, weekly), branch permissions, and task management. Cloud tasks require no machine to be on and are available to Pro, Max, Team, and Enterprise users.

Changed documents

Claude Code on the web [Source]

• Added a new "Schedule recurring tasks" section linking to the cloud scheduled tasks guide. [line ~173]

Commands [Source]

• Added /schedule [description] command for creating, updating, listing, or running cloud scheduled tasks conversationally. [line ~59]

Common workflows [Source]

• Added a new "Run Claude on a schedule" workflow section with a comparison table of cloud tasks, Desktop tasks, GitHub Actions, and /loop, plus guidance on writing effective prompts for autonomous runs. [line ~1290]

Desktop [Source]

• Added Computer Use feature section: lets Claude open apps and control your screen on macOS (Pro/Max only, research preview). Covers enabling it, required macOS permissions (Accessibility + Screen Recording), per-app permission tiers (view-only for browsers, click-only for terminals/IDEs, full control for everything else), and denied-apps configuration. [line ~106]
• Added Dispatch integration section: Dispatch (in the Cowork tab) can spawn Desktop Code sessions from a phone message; sessions appear with a "Dispatch" badge in the sidebar and send push notifications on completion. [line ~199]
• Added a scheduling options comparison table showing cloud vs. Desktop vs. /loop across dimensions like machine requirements, local file access, MCP servers, and minimum interval. [line ~376]
• Clarified that the Schedule page now supports both local and remote task types; creating a local task now requires choosing "New local task" from the modal. [line ~406]
• Updated the CLI comparison and capability tables to include Computer Use and Dispatch as Desktop-only features. [line ~546]

Index [Source]

• Added a "Schedule recurring tasks" section describing cloud tasks, Desktop tasks, and /loop. [line ~194]
• Added Dispatch to the "Work from anywhere" section and a "Run Claude on a recurring schedule" row to the platform comparison table. [line ~206]

MCP [Source]

• MCP server listing reordered (no new servers, ordering changes only).

Remote control [Source]

• Added a "Choose the right approach" comparison table covering Dispatch, Remote Control, Channels, Slack, and Scheduled Tasks — with trigger, runtime, setup, and best-for columns. [line ~186]

Sandboxing [Source]

• Added a "What sandboxing does not cover" section noting that built-in file tools (Read/Edit/Write) bypass the sandbox and use permissions directly, and that computer use on Desktop runs on the actual desktop rather than in an isolated environment. [line ~317]

Scheduled tasks [Source]

• Added a scheduling options comparison table (Cloud vs. Desktop vs. /loop) at the top of the document. [line ~6]
• Updated references to point to cloud scheduled tasks alongside Desktop tasks as the recommended durable scheduling alternative to session-scoped /loop. [line ~143]

API changes

New Documents

Scale to many tools with tool search [Source]

New dedicated guide for MCP tool search. Tool search is now on by default — tool definitions are withheld from context and loaded on demand. Supports up to 10,000 tools, returns 3-5 most relevant per search. Requires Claude Sonnet 4 or Opus 4 (no Haiku). Documents the ENABLE_TOOL_SEARCH environment variable (true/auto/auto:N/false), optimization tips for tool naming and descriptions, and limits.

Changed documents

Custom tools [Source]

• Major rewrite with improved structure: added a quick reference table at the top, separated tool definition from calling/registration, added Python examples alongside TypeScript throughout, and clarified handler return value requirements. [line ~1]
• Added guidance on making parameters optional in both Python and TypeScript. [line ~62]

MCP (Agent SDK) [Source]

• Removed the "Alternative: Change the permission mode" section and replaced it with a note clarifying that acceptEdits does not auto-approve MCP tools (only file edits), and that bypassPermissions is broader than necessary — recommending allowedTools wildcards instead. [line ~111]
• Tool search configuration details moved to the new dedicated tool-search.md page. [line ~199]
• Fixed a code example bug: changed message.content to message.message.content when iterating assistant message blocks. [line ~329]

Python SDK reference [Source]

• Added a full ToolAnnotations reference table (readOnlyHint, destructiveHint, idempotentHint, openWorldHint, title) with defaults and a code example. [line ~150]
• Added a @dataclass vs TypedDict runtime behavior note clarifying attribute access vs key access, with a ThinkingConfigEnabled example. [line ~544]
• Added max_output_tokens as a new AssistantMessageError literal type. [line ~1213]
• Noted that TypeScript supports additional hook events not yet in Python: SessionStart, SessionEnd, Setup, TeammateIdle, TaskCompleted, ConfigChange, WorktreeCreate, WorktreeRemove. [line ~1510]
• Improved the interrupt example: added draining of the interrupted task's ResultMessage before sending a new query, and added a note explaining buffer behavior after interrupt(). [line ~477]

Remote MCP servers [Source]

• Added new MCP servers: Google Cloud BigQuery, Zapier, Windsor.ai, Jotform, Aiwyn Tax, Brex, G2, Base44, Intuit TurboTax, Metaview. Various reorderings of existing entries.

TypeScript SDK reference [Source]

• Added a full ToolAnnotations reference table with defaults and a code example. [line ~62]
• Added SDKLocalCommandOutputMessage type (emitted for local slash commands like /voice or /cost). [line ~2517]
• Added fast_mode_state field ("off" | "cooldown" | "on") to SDKControlInitializeResponse. [line ~319]
• Added max_output_tokens as a new SDKAssistantMessageError literal type. [line ~680]
• supportedAgents() return type now documented as AgentInfo[]. [line ~297]

User input [Source]

• canUseTool callback now documented as receiving three arguments: toolName, input, and a new options/context parameter containing permission suggestions and a cancellation signal. [line ~35]
• Updated code examples to use isinstance(message, ResultMessage) and message.subtype == "success" instead of hasattr(message, "result"). [line ~113]