Claude Report — 2026-08-03
- No official Anthropic news, Claude Code changelog, or API changelog entries fall inside this window — CLI stays parked at v2.1.220 for a 9th straight day.
- All five tracked Agent repos and all five tracked MCP repos show zero activity in the window — the quietest tracking cycle so far.
- <a href="https://github.com/agentskillexchange/skills/commit/2966ec6">agentskillexchange/skills</a> lands one genuine human-authored skill (BeachFinder) amid its usual bot-churn commits.
- Claude Code lead Boris Cherny's 2+-week effort to have Claude rewrite the Claude Mac app in Swift draws public skepticism from John Gruber (61-point HN thread).
- A GitHub issue with reproducible benchmarks documents a Claude Gen-5 "nonsense detection" regression and undisclosed Fable-5→Opus-4.8 rerouting.
🚀 Anthropic Official
No relevant news today. Claude Code's CLI remains at v2.1.220 (released Jul 25) with no new release through Aug 3, and neither the Claude Platform API changelog nor anthropic.com/news lists any entry dated Jul 31 – Aug 3 — the most recent items there (Cognizant partnership, the open-weights position paper, the cybersecurity-incidents disclosure) all landed Jul 27–30, just outside this window.
🔌 Claude Code Plugins
Plugin install-rank cache is unavailable this cycle — top-10 plugin tracking is skipped per this run's instructions.
🛠️ Skills
agentskillexchange/skills
Merges a human-authored "Add BeachFinder Swim Spot Finder skill" (Integrations & Connectors, Aug 1) alongside its usual multiple-times-daily "regenerate from live data" / "sync published skills" bot cycles.
Unchanged in window: JSONbored/awesome-claude (latest fixes still dated Jul 30-31, already covered), Piebald-AI/claude-code-system-prompts (no new CLI release to diff), alirezarezvani/claude-skills (last Jul 17), anthropics/skills (last Jul 24).
🤖 Agents & Subagents
None of the top 5 updated in window.
Unchanged in window: nicobailon/pi-subagents (v0.40.0/v0.39.0 both landed Aug 1, already covered, nothing since), VoltAgent/awesome-claude-code-subagents (last change a Jul 31 README tweak, already covered), wshobson/agents (last Jul 18), anthropics/claude-agent-sdk-typescript (last release Jul 25), 0xmmo/crew (last Jul 4).
🔗 MCPs & Integrations
None of the top 5 updated in window.
Unchanged in window: awslabs/mcp (last commits Jul 30-31, already covered, none Aug 1-3), DeusData/codebase-memory-mcp (same, last activity Jul 31), shlokkhemani/rabbithole (last Jul 29), modelcontextprotocol/python-sdk (last Jul 29), modelcontextprotocol/typescript-sdk (last Jul 27).
💡 Community — Workflows & Ideas
Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App
John Gruber (Aug 2) recounts Claude Code lead Boris Cherny's two-plus-week, still-unfinished experiment having Claude rewrite the Claude Mac app pixel-by-pixel from Electron to Swift, and argues the exercise fixes stack problems but not the app's underlying design problems — a cautionary data point on agentic coding's limits at real-app scale.
Measurable quality regression in Claude generation 5
A reproducible-benchmark issue (Aug 3, using the open BullshitBench dataset) shows Sonnet 5/Opus 5/Fable 5 detect incoherent "nonsense" prompts 35 points worse than Gen 4.6/4.8 (0.87 → 0.52), emit 84-107% more tokens on the same zero-effort prompts, and shows Fable 5 silently rerouting to Opus 4.8 mid-session without disclosure.
Show HN: Gitleaks for Codex and Claude Code
Hamza (Aug 3) is an egress proxy that sits in front of Claude Code and Codex, masking detected cloud credentials, source-control tokens, and PII with anonymized placeholders before prompts leave the machine, addressing accidental secret leakage during agentic debugging.
Show HN: Changed how I use agent harnesses
Bifrost (Aug 3) is a self-hosted AI gateway offering one OpenAI-compatible API across 23+ providers (including Anthropic), with automatic failover, semantic caching, and MCP tool integration at roughly 11µs added overhead per request.
Gauge – see where your Claude Code subscription goes
Gauge (Aug 2) is a local, zero-dependency meter that reads Claude Code and Codex session transcripts directly to break down flat-rate subscription spend by 5-hour window, day, model, and project — without sending data externally.
📰 Quick Mentions
- Anthropic and OpenAI are competing to see whose agents can go rogue harder — Jul 31: frames both companies' recent agent-hacking disclosures as an escalating, market-driven "rogue agent" rivalry rather than isolated incidents.