Feature request: '/mcp reconnect' command to re-establish disconnected MCP servers without restarting the session

Feature request: "Reconnect all" / "Restart all MCP servers" command

Summary

When MCP servers disconnect mid-session (e.g. a network blip, a server-side restart, transient auth failures), there is currently no first-class way to ask Claude Code to attempt to reconnect them. The user has to either wait for an automatic reconnect to happen, restart the entire CLI session (losing conversation context), or manually figure out which server's process to nudge.

A /mcp reconnect (or similar) command that attempts to re-establish all currently-disconnected MCP servers in one action would be valuable.

Concrete scenario

In a long-running session, several MCP servers (e.g. atlassian-mcp, gitlab-mcp, synapse, slack-mcp) all disconnect simultaneously — likely because the host machine's network briefly dropped. The session inserts a system reminder noting which tools are no longer available.

The user would like to recover from this without restarting the session, because:

• The conversation thread contains substantial context (planning artifacts, intermediate findings).
• They've already authenticated to those services in this CLI session.
• Restarting would lose /resume-able state, transient task state, and any background processes the assistant has spawned.

Today the assistant's options are:

• Manually run claude mcp restart <name> from a separate terminal for each disconnected server (if such a command exists; the CLI's MCP subcommands aren't well-documented for this).
• Tell the user "the MCP servers are down, please reconnect them yourself" — which puts the burden on the user.
• Continue without those tools and hope they come back.

Proposed UX

Add a slash command (or expose via the existing /mcp menu):

• /mcp reconnect — attempt to reconnect all currently-disconnected servers.
• /mcp reconnect <name> — attempt to reconnect just the named server.
• /mcp restart — for currently-connected servers, restart them in place.

The command should:

1. Print which servers it's attempting to reconnect.
2. Show success/failure per server.
3. After completion, surface the deferred-tool list for newly-reconnected servers (so the assistant knows tools are available again without re-prompting).

Why this matters

For agentic workflows that stretch across many minutes (data engineering, multi-system orchestration, smoke tests), losing an MCP mid-session is common. The current recovery path is high-friction enough that users sometimes restart the session entirely — losing conversation state.

A one-liner reconnect command would make MCP disconnects a non-event.

---
Filed on behalf of a long-running session that lost atlassian-mcp, gitlab-mcp, slack-mcp, and several others mid-task. Manually reconnecting each was non-trivial.