Claude-Code v2.1.76 adds MCP elicitation, worktree optimizations, and numerous fixes

✍️ OpenClawRadar📅 Published: March 14, 2026🔗 Source
Claude-Code v2.1.76 adds MCP elicitation, worktree optimizations, and numerous fixes
Ad

What's new in Claude-Code v2.1.76

Claude-Code v2.1.76 is a significant update that adds new capabilities for MCP servers, improves monorepo workflows, and addresses numerous bugs affecting tool usage, session management, and user experience.

Key additions

  • MCP elicitation support: MCP servers can now request structured input during tasks via interactive dialogs using form fields or browser URLs
  • New hooks: Added Elicitation and ElicitationResult hooks to intercept and override responses before they're sent back to servers
  • Session naming: Added -n / --name <name> CLI flag to set a display name for sessions at startup
  • Monorepo optimization: Added worktree.sparsePaths setting for claude --worktree to check out only needed directories via git sparse-checkout in large monorepos
  • Compaction hook: Added PostCompact hook that fires after compaction completes
  • Effort control: Added /effort slash command to set model effort level
  • Quality surveys: Added session quality survey system where enterprise admins can configure sample rate via feedbackSurveyRate setting

Notable fixes

  • Deferred tools: Fixed deferred tools (loaded via ToolSearch) losing input schemas after conversation compaction, which caused array and number parameters to be rejected with type errors
  • Slash commands: Fixed showing "Unknown skill" and not finding soft-hidden commands when typing exact names
  • Plan mode: Fixed asking for re-approval after plan was already accepted
  • Voice mode: Fixed swallowing keypresses while permission dialog or plan editor was open, and fixed /voice not working on Windows when installed via npm
  • Context limits: Fixed spurious "Context limit reached" when invoking a skill with model: frontmatter on a 1M-context session
  • Model errors: Fixed "adaptive thinking is not supported on this model" error when using non-standard model strings
  • Bash permissions: Fixed Bash(cmd:*) permission rules not matching when quoted arguments contain #, and fixed "don't ask again" dialog showing full raw command for pipes and compound commands
  • Auto-compaction: Fixed retrying indefinitely after consecutive failures — now stops after 3 attempts with a circuit breaker
  • MCP: Fixed reconnect spinner persisting after successful reconnection
  • LSP plugins: Fixed not registering servers when LSP Manager initialized before marketplaces were reconciled
  • Clipboard: Fixed copying in tmux over SSH — now attempts both direct terminal write and tmux clipboard integration
  • Export: Fixed /export showing only filename instead of full file path in success message
  • UI issues: Fixed transcript not auto-scrolling to new messages after selecting text, and Escape key not working to exit login method selection screen
  • Remote Control: Fixed sessions silently dying when server reaps idle environment, rapid messages being queued one-at-a-time instead of batched, stale work items causing redelivery after JWT refresh, and bridge sessions failing to recover after extended WebSocket disconnects
Ad

Performance and UX improvements

  • Worktree startup: Improved --worktree startup performance by reading git refs directly and skipping redundant git fetch when remote branch is already available locally
  • Background agents: Killing a background agent now preserves its partial results in conversation context
  • Model fallback: Notifications are now always visible instead of hidden behind verbose mode, with human-friendly model names
  • Terminal themes: Improved blockquote readability on dark themes — text is now italic with left bar instead of dim
  • Worktree cleanup: Worktrees left behind after interrupted parallel runs are now automatically cleaned up
  • Session titles: Remote Control session titles now derived from first prompt instead of showing "Interactive session"
  • Voice improvements: /voice now shows dictation language on enable and warns when language setting isn't supported for voice input
  • Plugin directories: Updated --plugin-dir to only accept one path to support subcommands — use repeated --plugin-dir for multiple directories
  • VSCode fix: Fixed gitignore patterns containing commas silently excluding entire filetypes from the @-mention file picker

This release focuses on improving MCP server interactions, fixing persistent tool and session management issues, and enhancing the developer experience across terminal, voice, and remote workflows.

📖 Read the full source: GitHub Claude-Code

Ad

👀 See Also

Fixing OpenClaw Browser CAPTCHAs with Camoufox and CLI Wrapper
Tools

Fixing OpenClaw Browser CAPTCHAs with Camoufox and CLI Wrapper

OpenClaw's built-in Chromium browser triggers bot detection through Chrome DevTools Protocol, JavaScript injection artifacts, and hardware fingerprinting inconsistencies. The solution uses Camoufox (a Firefox fork) modified at the C++ level and wrapped in a CLI that returns accessibility-tree snapshots to reduce token usage.

OpenClawRadar
Engram: Open-source memory layer for Claude Code and MCP clients
Tools

Engram: Open-source memory layer for Claude Code and MCP clients

Engram is an open-source memory layer that works as an MCP server with any client like Claude Code, Cursor, or Windsurf. It stores unlimited memories with semantic vector search, achieves 80% accuracy on LOCOMO benchmark, and uses about 800 tokens per query versus 5K+ for file-based approaches.

OpenClawRadar
Claudetop: Real-Time Cost Monitoring for Claude Code Sessions
Tools

Claudetop: Real-Time Cost Monitoring for Claude Code Sessions

Claudetop is an htop-like tool that shows real-time spending, cache efficiency, and model comparisons for Claude Code sessions. It provides slash commands like /claudetop:stats and smart alerts for cost milestones and efficiency issues.

OpenClawRadar
Scaling Karpathy's Autoresearch with 16 GPUs: Results and Methods
Tools

Scaling Karpathy's Autoresearch with 16 GPUs: Results and Methods

The SkyPilot team gave Claude Code access to 16 GPUs on a Kubernetes cluster to run Karpathy's Autoresearch project. Over 8 hours, the agent submitted ~910 experiments, reduced validation bits per byte from 1.003 to 0.974 (2.87% improvement), and reached the best validation loss 9x faster than sequential execution.

OpenClawRadar