Anthropic News Flash | Claude Code v2.1.145 and MCP Tunnels

[1] Claude Code v2.1.145 Released — Permission Prompt Bypass Bug Fixed and claude agents --json Added

We are reporting that Claude Code v2.1.145 was published on May 19, 2026 at 21:31 UTC (primary sources: https://github.com/anthropics/claude-code/releases/tag/v2.1.145, https://code.claude.com/docs/en/changelog). This release focuses on enhanced multi-agent observability and includes several new features along with an important security fix.

Impact for engineers: ★★★

The biggest highlight is the fix for a bug where bare variable assignments to environment variables inside Bash commands were auto-approving permission prompts even when outside the allowlist. Given the broad impact, an immediate update is recommended as a mandatory security measure. The newly added claude agents --json command lists running Claude sessions as JSON, enabling direct integration with tmux-resurrect, status bar tools, and custom dashboards.

On the OTEL side, agent_id / parent_agent_id were added to claude_code.tool spans, and background sub-agent spans are now correctly nested under their parent Agent tool spans. This significantly increases the value of building out a multi-agent tracing infrastructure. Status line JSON input now includes GitHub repository and PR information when detected, and the /plugin Discover/Browse screens now allow you to review plugin details (commands, agents, skills, hooks, MCP/LSP servers) before installation.

Other notable changes include displaying the number of agents awaiting input in the claude agents terminal tab title, adding background_tasks / session_crons fields to Stop and SubagentStop hook inputs, mouse support in slash command and @mention suggestion lists in fullscreen mode, a fix for MCP prompt slash commands not displaying required argument names, and resolution of an issue where Agent Teams with non-ASCII names caused API errors.

[2] Four Platform Docs Updates on 2026-05-19 — MCP Tunnels Preview and Self-Hosted Sandbox

We are reporting that four new features were added simultaneously to the Platform Docs release notes on May 19, 2026 (primary source: https://platform.claude.com/docs/en/release-notes/overview). The biggest highlight is the MCP Tunnels Research Preview, which enables connections to MCP servers inside private networks.

Impact for engineers: ★★★

MCP Tunnels is a feature that removes a major barrier to connecting MCP servers in on-premises or private VPC environments to Claude, making it a long-awaited update in enterprise adoption scenarios. On the same date, Managed Agents gained a self-hosted Sandbox, enabling tool execution outside Anthropic infrastructure. This is a feature that could become a prerequisite for enterprise adoption in regulated industries and organizations with data sovereignty requirements, and can be immediately used as a deciding factor in internal compliance reviews.

The third addition is live updates to MCP server/tool configuration during active sessions, allowing long-running agents to switch connection targets without being stopped. The fourth is automatic output spillover: when the output of agent_toolset or MCP tools exceeds 100K tokens, it is automatically saved to a sandbox file. The model receives a truncated preview with the file path and can retrieve the full content as needed, preventing context exhaustion and improving stability for large-scale agent tasks.

[3] "Widening the conversation on frontier AI" — AI Alignment Dialogue with 15+ Religious and Philosophical Organizations

We are reporting that "Widening the conversation on frontier AI" was published on Anthropic News on May 19, 2026 (primary source: https://www.anthropic.com/news/widening-conversation-ai). The article announces that Anthropic is advancing research on the "moral formation" of AI systems through dialogue with more than 15 organizations and groups representing religious, philosophical, and cultural traditions.

Impact for engineers: ★★

This initiative incorporates diverse perspectives into shaping the values reflected in Claude, and it makes explicit a direction of embedding cultural and ethical frameworks into the design stage — not just technical optimization — as an approach to AI alignment. While there are no direct API changes, it is worth tracking for understanding long-term shifts in Claude's response tendencies. In particular, this official communication can serve as reference material for ethics reviews and compliance processes when deploying Claude in enterprise settings, and organizations drafting AI governance documents would be wise to keep it as a potential citation.

[4] Code w/ Claude London Opens — All Sessions Streamed Live, Tokyo Continues June 5–6

We are reporting that Code w/ Claude London, Anthropic's developer conference, opened on May 20, 2026 (local time) (primary sources: https://www.anthropic.com/code-with-claude, https://www.anthropic.com/news). The Day 1 keynote and all breakout sessions were streamed live, allowing engineers unable to attend in person to follow the content in real time.

Impact for engineers: ★★

Announcements from London (new features, new APIs, case studies, etc.) are likely to generate breaking news between May 20–21, requiring follow-up in subsequent collection reports. Registration for the Tokyo venue (June 5–6) is also still open, making it the largest official Anthropic event for engineers in Japan. Readers who have not yet registered for Tokyo are encouraged to check the registration page promptly. The Clauder Navi editorial team plans to publish separate articles covering London session highlights and on-site reports from Tokyo.

[Next Steps] Immediate v2.1.145 Update and Internal MCP Tunnels Testing

In the editorial team's assessment, the top priority task to tackle in the near term is the immediate update to Claude Code v2.1.145. The permission prompt bypass bug is a mandatory security fix, and prompt deployment across both CI and developer environments is advisable. In parallel, now is the right time to begin internal testing of the MCP Tunnels Research Preview: organizations with scenarios for connecting Claude to on-premises databases or internal APIs should start designing PoC architectures and evaluating connection paths to gain a first-mover advantage. The self-hosted Sandbox provides grounds for adoption decisions in regulated industries and organizations with data sovereignty requirements, so sharing this with compliance teams simultaneously is a prudent course of action. We will continue to track announcements from Code w/ Claude London and Tokyo in this news section.

[Sources] Primary Information Links Used in This Article

• Claude Code Releases (v2.1.145): https://github.com/anthropics/claude-code/releases/tag/v2.1.145
• Claude Code Changelog: https://code.claude.com/docs/en/changelog
• Platform Docs Release Notes: https://platform.claude.com/docs/en/release-notes/overview
• Anthropic News (Widening the conversation on frontier AI): https://www.anthropic.com/news/widening-conversation-ai
• Code w/ Claude Official: https://www.anthropic.com/code-with-claude
• Anthropic News (Top): https://www.anthropic.com/news