Daily AI Builder Brief - April 21, 2026

Executive Summary

Claude Code 2.1.116 shipped April 20 with up to 67% faster /resume on 40MB+ sessions and quicker MCP stdio startup.
A sandbox auto-allow bug that could let rm or rmdir hit /, $HOME, or critical paths is fixed; upgrading isn't optional.
Amazon committed up to $25B more to Anthropic: $5B now, $20B milestone-gated, on top of $8B prior. Initial round values Anthropic at $380B.
Anthropic secured up to 5 gigawatts of capacity and will bring ~1GW of Trainium2/3 online by year-end; Claude throughput should climb through 2026.
MCP's draft authorization spec now codifies OAuth 2.1 with PKCE, resource indicators, and step-up (incremental) scope consent.
Codex's April 20 update added MCP Apps tool calls and namespaced MCP registration, so cross-client MCP UX is converging fast.

Claude Code

2.1.116: faster `/resume`, lighter MCP startup, sandbox `rm` hole closed

/resume is up to 67% faster on 40MB+ transcripts and handles dead-fork entries cleanly. Multi-stdio MCP startup is quicker since resources/templates/list now defers to the first @-mention. The security fix matters: sandbox auto-allow no longer bypasses the dangerous-path check for rm/rmdir against /, $HOME, or critical directories. Agent hooks: now fire when running as the main-thread agent via --agent, and /doctor can open mid-turn. Claude Code changelog

Anthropic

Amazon expands the deal: up to $25B, 5GW of compute, Trainium-heavy

Amazon will invest up to $25B more in Anthropic on top of the existing $8B: $5B immediate, $20B tied to commercial milestones. Anthropic commits $100B+ to AWS over 10 years and locks in up to 5 gigawatts of capacity, with ~1GW of Trainium2 and Trainium3 online by year-end. Initial round is at a $380B valuation. For builders: expect steadier Claude availability and throughput across the year. Anthropic

MCP Ecosystem

Authorization spec: OAuth 2.1 + incremental scope consent

The draft MCP authorization spec now codifies OAuth 2.1 with PKCE, resource indicators, and step-up authorization. Servers expose a minimal scopes_supported set; clients request more via scope-challenge handling only when a specific tool needs it. If you ship MCP servers, design for per-tool scopes and incremental consent UX, not one upfront grant. MCP auth spec

Codex adopts MCP Apps and namespaced server registration

OpenAI's Codex April 20 drop expanded MCP support with MCP Apps tool calls, namespaced server registration, parallel-call opt-in, and sandbox-state metadata. Same-day cadence with Claude Code 2.1.116's MCP work means MCP UX patterns are crystallizing across clients; pick server names you can live with. Codex notes via Releasebot

Broader AI

Nothing today.

What This Means For Builders

Upgrade Claude Code to 2.1.116 now. The sandbox rm fix alone makes this mandatory.
Architect MCP servers around per-tool scopes and step-up consent; big upfront scopes will feel broken under the new spec.
You can plan agent workloads assuming Claude throughput grows, not shrinks, through 2026.
Pick stable, unique MCP server names; namespaced registration is now showing up in multiple clients.