Daily AI Builder Brief - April 18, 2026

Executive Summary

Claude Code 2.1.113 swaps the bundled JS CLI for a native per-platform binary, which changes install expectations on CI runners.
New sandbox.network.deniedDomains lets deny rules win over wildcard allow rules, closing a real exfil gap.
Bash permission rules now match commands wrapped in env, sudo, watch, ionice, setsid; prior allowlists may have been wider than you thought.
Bash(find:*) allow rules no longer auto-approve find -exec or -delete; audit your agent allowlists.
Claude Opus 4.7 is live with a new xhigh effort between high and max, plus task budgets that give agents a running token countdown.
/ultrareview ships for parallel multi-agent code review, and /less-permission-prompts scans transcripts and proposes allowlists.
Anthropic Labs launched Claude Design for prototypes, slides, and one-pagers built on Opus 4.7, with export to Canva, HTML, and Claude Code.
MCP upstream is quiet: no new spec or official-server release in the 48 hour window.

Claude Code

2.1.113 (Apr 17): native binary, sandbox hardening, MCP watchdog fix

Native binary spawn replaces the bundled JS entry. Added sandbox.network.deniedDomains. /private/{etc,var,tmp,home} are now dangerous under Bash(rm:*). Fixed MCP concurrent-call timeout where one tool's message could disarm another's watchdog. Fixed dangerouslyDisableSandbox silently skipping prompts. (changelog)

2.1.111 (Apr 16): Opus 4.7, /ultrareview, /effort slider

Opus 4.7 xhigh with /effort arrow-key slider. Auto mode on Opus 4.7 for Max subscribers. /ultrareview runs parallel cloud review with a diffstat in the launch dialog. New /less-permission-prompts skill. Ctrl+A/Ctrl+E now match readline on multiline input. (changelog)

Anthropic

Claude Opus 4.7 (Apr 16)

New xhigh effort sits between high and max. Opus 4.7 at xhigh with 100k tokens beats 4.6 at max with 200k on Anthropic's internal scoring. Task budgets give the model a running countdown so it finishes gracefully. 1M context, 128k output, adaptive thinking. (post)

Claude Design (Apr 17)

Anthropic Labs product for prototypes, slides, and one-pagers. Reads your codebase and design files to apply your design system. Exports to .zip, .pdf, .pptx, Canva, HTML, and Claude Code. Built on Opus 4.7 vision. (TechCrunch)

MCP Ecosystem

Nothing in the 48 hour window

No new spec cut, no new official server. The most recent upstream post remains the Apr 8 maintainer expansion. (blog)

Broader AI

Nothing today.

What This Means For Builders

Audit your Claude Code permission config today: wrapped-command coverage and find tightening can break agents that relied on loose allowlists.
If you run agents headless in CI, pin the 2.1.113 binary path and test the native-spawn behavior before rolling forward.
Try xhigh for long coding loops before reaching for max: cheaper tokens, similar scores.
Task budgets are worth wiring into your agent harness so runs end cleanly instead of mid-thought.