Claude Code 2.1.186 adds claude mcp login and auto-replies after ! shell commands

• ClaudeCodeLog's release summary says Claude Code 2.1.186 ships 33 CLI changes, led by CLI-based MCP auth, automatic assistant replies after ! shell commands, and stricter named-subagent permission enforcement.
• According to the 2.1.186 changelog, claude mcp login <name> and claude mcp logout <name> work without opening /mcp, and --no-browser supports stdin redirect flows over SSH.
• The changelog also flips ! shell commands into an active turn boundary: Claude now responds to command output automatically unless respondToBashCommands is set to false.
• The same release notes fix a real policy hole, because named subagent spawns had not been enforcing Agent(type) deny rules and Agent(x,y) allowed-types restrictions.
• ClaudeCodeLog's diff summary adds one more small but useful detail: the release also introduces --no-browser, removes CLAUDE_CODE_CONNECT_TIMEOUT_MS, and lands just 2 days and 10 hours after 2.1.185.

You can jump straight to the source changelog, compare it with the tiny 2.1.185 release, and inspect the diff metadata that tracks bundle size, removed surfaces, and prompt-token stability.

The most practical addition is a CLI path for MCP authentication. The 2.1.186 changelog adds claude mcp login <name> and claude mcp logout <name>, which means remote or headless setups no longer have to bounce through the interactive /mcp menu.

The SSH detail matters more than the headline. As the changelog entry notes, --no-browser supports stdin redirect completion, and ClaudeCodeLog's surface diff separately lists --no-browser as a newly added option.

Claude Code now treats ! shell commands more like a conversational handoff than a passive context dump. ClaudeCodeLog's highlight post says command output now triggers an immediate assistant response.

The changelog spells out the rollback switch: set "respondToBashCommands": false in settings.json to keep the previous context-only behavior. That makes this a default workflow change, not just a new flag hidden in docs.

One of the sharper fixes is buried in the permissions layer. The changelog says named subagent spawns were failing to enforce both Agent(type) deny rules and Agent(x,y) allowed-types restrictions.

That same section adds a second guardrail. According to the security and safety note, background subagents now surface permission prompts in the main session instead of auto-denying them, and the dialog shows which agent requested the tool.

A lot of the remaining 30 changes are quality-of-life cleanup for people already deep in Claude Code's agent UI. The release notes add status filtering with f in /workflows, add a Skills section to /plugin Installed, and add teammateMode: "iterm2" with a warning when auto mode cannot find the it2 CLI.

The same changelog also fixes several agent-workflow edge cases:

• Background agents no longer leave Esc and Ctrl+C unresponsive after the main turn ends.
• Teammates spawned via tmux or pane backends now inherit the leader's --effort level.
• agent({schema}) workflow subagents now abort after 5 repeated schema-validation failures instead of looping forever.
• The agent now gets reminded to compact its MEMORY.md index when it nears the size limit.

ClaudeCodeLog's additional modifications pegs 2.1.186 at 2 days, 10 hours, 4 minutes after 2.1.185, which itself was basically a one-line stream-stall tweak in the previous release summary. This one is a much fatter patch by comparison.

The same diff snapshot in ClaudeCodeLog's metadata post says bundle size dropped by 254.1 kB, prompt files and prompt-token counts stayed flat, CLAUDE_CODE_CONNECT_TIMEOUT_MS was removed, and the claude-inv- model surface disappeared from the CLI listing.