The AI Catchup -- May 19, 2026

Welcome back. Some weeks are broad; this one is deep. If you use Claude Code's agent view or run remote MCP servers, there is one release you should not miss.

Let us get into it.

This Week in AI

Anthropic shipped Claude Code 2.1.142 (vendor release May 14). The headline feature is per-dispatch configurability for background sessions: eight new claude agents flags that let each dispatched run override directories, settings, MCP config, plugin directories, permission mode, model, effort, and the permission bypass.

Why it matters: if you are using claude agents as a real dashboard (not just a demo), you eventually want heterogeneous runs. One agent should run tests at low effort, another should do a high-effort migration, a third might need a different MCP config or plugin set. Until now, your options were daemon-level defaults, wrapper scripts, or multiple installs. 2.1.142 makes the dispatch surface match the way teams actually use agent dashboards.

Tool Spotlight

The other quietly critical fix in 2.1.142 is that MCP_TOOL_TIMEOUT is now honored end to end for remote HTTP and SSE MCP servers. Previously, long-running tool calls could still die at 60 seconds because the underlying fetch timeout was capped regardless of your configured value.

If you run remote MCP servers with tools that legitimately take more than a minute (code execution, data processing, long network calls), this is the update that stops "it always times out at 60s" from being a weekly mystery.

What We Are Watching

Model defaults are also shifting. Fast mode now defaults to Opus 4.7 instead of Opus 4.6, with an explicit escape hatch: CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1.

If you have any brittle prompts, tool-call expectations, or evaluations that were tuned on 4.6, set the override and upgrade anyway. You can then migrate on your timeline instead of during an incident.

Comparison Reading

If you have not read it recently, the Codex CLI vs Claude Code vs Cursor architecture deep-dive is still the right context for why these fixes matter. Claude Code's MCP-first design is what makes a single bug in remote HTTP/SSE tool timeouts so operationally painful.

Quick Hits

• New claude agents flag set: --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, --dangerously-skip-permissions.
• Background-session stability: multiple agent-view and daemon fixes shipped alongside the main release (worktree recognition, sleep/wake resilience, and cleaner upgrades).

---

That is it for this week. The AI Catchup publishes every Tuesday. If you found this useful, subscribe to get it in your inbox.

Until next week -- stay caught up.