Claude Code adds usage insights, keyboard diff scrolling, MCP tunnels research preview, and self-hosted sandboxes

Anthropic's Claude Code release is mostly developer-loop polish and enterprise plumbing, but two pieces matter. First, MCP tunnels (Research Preview) let Claude connect to MCP servers behind firewalls without opening inbound ports — a long-standing enterprise blocker for adopting MCP-based tools at scale. Second, self-hosted sandboxes for Claude Managed Agents let enterprises run agent execution inside their own boundary, addressing data-residency and audit concerns that have kept regulated industries cautious.

Mechanically the usage insights expose token consumption by session, command, and tool — directly relevant to the same Microsoft-style cost-blowup that just cost Anthropic seats. Better diff scrolling and Markdown task list rendering are small but cumulative quality-of-life wins that matter for long sessions. The enterprise MCP connector setting on Claude.ai centralizes governance over what cloud MCPs an org's users can install.

Competitive context: this lands the same week Microsoft canceled most internal Claude Code licenses over token costs and Boris Cherny was promoting auto mode on X. The enterprise features (tunnels, sandboxes, governance) are clearly designed to give the next 'Microsoft' a less escape-hatch-shaped reason to leave. Meanwhile a developer published a directory of 28,577 MCP servers, signaling that MCP discovery and trust are becoming first-order problems Anthropic will need to address.

What to watch: whether MCP tunnels move from Research Preview to GA before competitors ship equivalents, whether self-hosted sandboxes win back regulated-industry deployments, and whether Anthropic adds explicit cost-control primitives (per-team budgets, hard token caps) to the enterprise tier. The token-cost story isn't going away.