Claude Code 2.1.76 adds MCP elicitation, max effort, and transcript-off support

• Anthropic shipped Claude Code 2.1.76 with 35 CLI changes, led by MCP elicitation so servers can request structured input mid-task through forms or a browser URL, plus new Elicitation and ElicitationResult hooks to intercept replies before they are returned 2.1.76 release CLI changelog.
• Session control expanded across the CLI and Remote Control: /effort now supports a max mode that “reasons for longer,” claude --name can label sessions at startup, and Remote Control can start entirely new sessions instead of only attaching to existing ones max effort post Remote Control.
• Operational controls got tighter too: Claude Code now supports disabling transcript persistence entirely, adds a PostCompact hook after conversation compaction, and stops auto-compaction after three consecutive failures instead of retrying forever further changes 2.1.76 release.
• The preceding 2.1.75 release matters because it set up the same theme: Opus 4.6 now gets a default 1M context window on Max, Team, and Enterprise, permission denials ask users for clarification instead of guessing, and memory files now expose last-modified timestamps 2.1.75 release prompt updates.

The biggest implementation change is MCP elicitation. According to the CLI changelog, MCP servers can now pause a task and ask for “structured input mid-task” via form fields or a browser URL, which turns user clarification into a first-class protocol step instead of a free-form chat detour. The same release adds Elicitation and ElicitationResult hooks, so teams running custom MCP integrations can intercept or override those responses before they go back to the server CLI changelog.

Claude Code also added primitives for longer-running repo work. The changelog details say worktree.sparsePaths lets claude --worktree use git sparse-checkout for large monorepos, and PostCompact fires after compaction completes. In the same build, deferred tools loaded via ToolSearch no longer lose their schemas after compaction, fixing a failure mode where array and number parameters were rejected with type errors CLI changelog.

The operational fixes are aimed at sessions that run unattended for longer. Anthropic's release thread says auto-compaction now has a circuit breaker after three failed attempts, while Remote Control fixes address idle-session death, one-at-a-time queueing, stale work-item redelivery after JWT refresh, and bridge-session recovery after long WebSocket disconnects CLI changelog.

Anthropic also exposed more direct runtime controls. In an end-of-week product post, a team member said /effort can now be set to max, which “reasons for longer and uses as many tokens as needed,” but must be enabled per session because it burns through usage limits faster max effort post. That makes effort an explicit session-level tradeoff instead of a hidden model behavior.

Session management got broader across surfaces. Remote Control can now start new sessions, not just operate existing ones Remote Control. Voice mode is “rolled out to 100% of users,” including Claude Code Desktop voice rollout, and Claude Code on web and desktop can run a setup script setup script. The companion product thread adds smaller but practical controls: claude --name <NAME> labels a session from startup, /color changes the prompt input color, and sessions can be auto-named after plan mode other new things.

The lower-level 2.1.76 diff adds one more deployment-relevant knob. The further changes mention an option to “disable transcript persistence entirely,” alongside new config and env-surface additions and a warning when the current model is scheduled for retirement further changes. Anthropic also changed --plugin-dir to accept only one path per flag, requiring repeated flags for multiple directories CLI changelog.

Version 2.1.75 landed less than two days before 2.1.76 and established the same pattern of control and reliability work. The headline change was that Opus 4.6 now gets a default 1M context window on Max, Team, and Enterprise plans, where that capacity had previously required extra usage 2.1.75 release. The update also added memory-file timestamps so Claude can distinguish “fresh vs. stale” memories and show hook source in permission prompts, which matters when settings, plugins, or skills trigger confirmation flows 2.1.75 changelog.

The system prompt changed in ways engineers will feel during tool use. The prompt diff says Bash guidance now more forcefully bans sleep for long-running commands and pushes run_in_background instead, while denied tool permissions now prompt for a reason when user intent is unclear rather than guessing the next step. The same prompt diff removed the instruction that Read truncates lines over 2000 characters, changing how the agent should interpret missing content in long single-line files prompt updates.

There were also fixes directly tied to long-context and enterprise use: token estimation no longer over-counts thinking and tool-use blocks, which had caused premature context compaction, macOS startup skips unnecessary subprocess spawns on non-MDM machines, and Windows managed settings now require the newer Program Files path after the old fallback was removed 2.1.75 changelog full 2.1.75 changelog.