Claude Code's summer drop: artifacts with MCP connectors, /fork, /teleport, and auto mode on three clouds

Claude Code's changelog between the end of July and August 11 reads like three releases stacked on top of each other, and if you only skim it you will miss the one that changes what the tool is for. Artifacts published from Claude Code now read live data through MCP connectors, per viewer. A dashboard you publish is not a snapshot of your data; it is a page that queries the viewer's own connectors when they open it. That moves Claude Code from a coding CLI toward a delivery platform, and the rest of the drop is mostly the plumbing to make that safe inside a company.

Late July: artifacts, /fork, and auto mode everywhere

• Live MCP connector data in published artifacts, resolved per viewer, plus public artifact sharing. The artifact is a client; the viewer's connectors are the data source.
• /fork to branch a session and /doctor to diagnose a broken setup.
• Editor roles on shared artifacts, so a published page can be co-owned rather than one person's file.
• Auto mode on AWS Bedrock, Google Cloud Agent Platform, and Microsoft Foundry, so the same autonomy level works behind whichever cloud your procurement picked.
• Improved subagent workflows and permissions, following the earlier July additions of Opus 5 as the default Opus, nested subagents, dynamic workflows, and admin console tabs for value and usage.

Underneath, Claude Code already supports the MCP spec dated 2026-07-28, the stateless request/response revision with OAuth/OIDC. That is what makes per-viewer connector reads workable: no long-lived session to pin an artifact to. If you have not followed that change, the MCP primer is the starting point and Claude Code with MCP covers the client side.

August 6: the boring release that matters

The August 6 batch was fixes across session handling, security, and reliability, better worktree support, and improved tool search on Google Vertex AI for Claude 4.5-generation models and newer. The same day, Claude for Enterprise got a beta that automatically scans third-party skills and plugins for malicious content at upload or edit time. If you ship skills or MCP servers to enterprise customers, expect your artifacts to be scanned before anyone runs them.

August 11: enterprise controls and /teleport

• owner/* marketplace controls for an entire GitHub organization, so an admin can allow or block plugin sources org-wide instead of per repo.
• /teleport to pull a cloud session down to your local machine and keep working with local tools.
• Better review commands and another round of security and session-resume fixes.
• The Compliance API extended in beta to Cowork and Claude Code across desktop, web, mobile, and CLI.

Then, later in August: GitLab merge request support inside --worktree and the agents view, forward_user_identity for per-user spend attribution when Claude Code runs behind a gateway, and an ANTHROPIC_DEFAULT_MODEL environment variable.

The feature list is long, but the thread through it is one question: can a company run this thing for two hundred people and still know who did what and what it cost.

What to change in your setup

• Turn on forward_user_identity if you front Claude Code with a gateway. Per-user attribution is the difference between a cost report and an argument.
• Set ANTHROPIC_DEFAULT_MODEL in CI and shared environments so a model bump is a config change, not a surprise in someone's session.
• Use worktrees for agent runs, and now for GitLab MRs too, so an autonomous branch never touches your checked-out tree.
• Audit marketplace sources at the org level before individual developers add plugins; the owner/* controls make that a policy, not a request.
• Treat published artifacts as applications. They read live connector data per viewer, so review what an artifact can reach the same way you would review an internal tool.

The honest gap

This is all changelog and release-note material; I have not benchmarked the Vertex tool search improvement or verified the per-viewer connector behavior against a hostile artifact. Auto mode being available on three clouds also does not mean the permission and audit surfaces are identical on each, and the notes do not say. Verify the cloud you actually use before you promise a client the same guardrails.