Claude docs changes for February 17th, 2026 [diff]

Executive Summary

Plugin marketplace documentation greatly expanded with new sections on plugin sources, strict mode behavior, and version resolution/release channels
Two new remote MCP servers added: Microsoft Learn and MSCI
Jailbreak mitigation docs updated: model recommendation changed to Claude Haiku 4.5 and structured outputs now recommended for classification responses
Claude Code version 2.1.43 fixes AWS auth refresh hang and structured-outputs header issues on Vertex/Bedrock

Fixed AWS auth refresh hanging indefinitely by adding a 3-minute timeout
Fixed spurious warnings for non-agent markdown files in .claude/agents/ directory
Fixed structured-outputs beta header being sent unconditionally on Vertex/Bedrock

Added Microsoft Learn as a new remote MCP server for searching trusted Microsoft documentation. [diff]

Updated product description: Claude Code is now described as an "AI-powered coding assistant that helps you build features, fix bugs, and automate development tasks" rather than "an agentic coding tool that reads your codebase, edits files, and runs commands."

Clarified strict field description: now explains that true makes plugin.json the authority for component definitions (with marketplace entry able to supplement), while false means the marketplace entry defines everything. [line 248] [Source]
Added new "Plugin sources" section with a reference table listing all supported source types (relative path, github, url, npm, pip) and their fields. [lines 262-281] [Source]
Added clarification distinguishing marketplace sources (where to fetch marketplace.json) from plugin sources (where to fetch individual plugins), with a concrete example. [lines 283-292] [Source]
Added a dedicated "Strict mode" section explaining the behavior of strict: true vs strict: false in detail, including when to use each. [lines 448-463] [Source]
Added new "Version resolution and release channels" section explaining how plugin versions determine cache paths and update detection, plus guidance on setting up stable/latest release channels via separate marketplaces. [lines 657-764] [Source]
Simplified external dependency guidance: removed "Option 2: Restructure your marketplace" approach, leaving only symlinks as the recommended solution for accessing files outside the plugin directory.

Clarified that marketplace plugins are installed to ~/.claude/plugins/cache and that --plugin-dir plugins are only available for the duration of a session. [lines 406-412] [Source]
Added note that version bumps in plugin.json are required for cached plugins to propagate updates to existing users, with guidance for marketplace-managed versions. [lines 772-773] [Source]

Updated obfuscated email link for contacting sales about full thinking output for Claude 4 models (no content change, URL encoding updated).

Updated obfuscated email link for contacting sales about full thinking output for Claude 4 models (no content change, URL encoding updated).

Updated obfuscated email link for contacting support about higher rate limits (no content change, URL encoding updated).

Removed claim that Claude is "far more resistant to jailbreaking than other major LLMs, thanks to Constitutional AI."
Updated harmlessness screen recommendation from Claude Haiku 3 to Claude Haiku 4.5, and added recommendation to use structured outputs to constrain classifier responses to a simple classification. [lines 7-8] [Source]

Added Microsoft Learn as a new remote MCP server for searching trusted Microsoft documentation (https://learn.microsoft.com/api/mcp). [lines 559-565] [Source]
Added MSCI as a new remote MCP server for querying MSCI indexes (https://mcp.msci.com/index/index-mcp/v1.0/mcp). [lines 599-605] [Source]