Claude Report — 2026-09-19

🚀 Anthropic Official

Claude Code v2.1.277 (Sep 18)

Adds AGENTS.md support — reads AGENTS.md instead of CLAUDE.md when no CLAUDE.md exists in a project (editable under "Project instructions" in /config; not yet on Bedrock, Vertex or Foundry). Strips invisible Unicode formatting/tag characters from prompts before review and frames subagent results so their content can't masquerade as top-level session instructions. Adds CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 and an optional headers: map for Claude apps gateway upstreams, plus 100+ fixes (hanging claude -p/Agent SDK sessions after internal errors, empty-text-block crashes, unexpected logout on mixed CLI versions, /plugin install breaking on reinstall, Grep/Glob false negatives under resource limits, Edit backslash/Unicode escaping).

Claude Code v2.1.278 (Sep 19)

Switches auto mode to default to the server-side classifier for Claude API/Enterprise users and on Bedrock, Vertex, Foundry and gateways, which no longer charges for classifier overhead; adds CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out and an Auto mode server row in /status, and warns on billed fallback.

claude-agent-sdk-typescript v0.3.277–v0.3.278 (Sep 18–19)

Adds an optional builtin field to SlashCommand, pasted_content on SDKUserMessage, remote-session latency fields on success results, and a 'userSettings' updateSettings() source (accepts effortLevel); fixes resumed/forked session cost totals, then syncs to Claude Code v2.1.278.

claude-agent-sdk-python v0.2.156 (Sep 18)

Updates the bundled Claude CLI to v2.1.276 — the Python SDK trails the TypeScript SDK and CLI by one to two releases this cycle.

Partnering with Accenture on embedded evaluation (Sep 18)

Anthropic and Accenture's Faculty unit partner on independent evaluation — red-teaming, alignment assessments and safeguard testing — with evaluators getting employee-level internal access; each side expects to invest at least $1B over five years, and Anthropic says it's in talks with METR and other nonprofit evaluators for further pilots.

Anthropic discloses a Bay Area wet biology lab (Sep 18)

Confirms to TechCrunch it runs a physical lab where AI models help run real biology experiments, distinct from its Sep 17 biomolecular-modeling software research; head of Life Sciences Eric Kauderer-Abrams says wet-lab work remains "the final test," and the disclosure sits alongside the April Coefficient Bio acquisition and the new Life Sciences Verification Program.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

cloudflare/security-audit-skill

Open-sources a coding-agent skill that runs six-phase security audits — recon, coverage-led hunting, candidate validation, structured output, independent re-verification, target-neutral reporting — as the single-repo seed of Cloudflare's internal vulnerability-discovery harness; surges to trending #1-adjacent with 9,000+ stars added since its Sep 17 coverage.

Unchanged in window: eugeniughelbur/obsidian-second-brain (last merge Sep 18 already reported, no new commits), tt-a1i/archify, ayghri/i-have-adhd, K-Dense-AI/scientific-agent-skills, affaan-m/ECC, agentskillexchange/skills.

🤖 Agents & Subagents

nicobailon/pi-subagents v0.69.0

Ships a typed acceptance-gates bridge that pipes JSON stdout into structured output for subagent verification (Sep 18).

Unchanged in window: VoltAgent/awesome-claude-code-subagents, anthropics/commerce-agents.

🔗 MCPs & Integrations

oraios/serena

Upgrades to MCP SDK v2.2.0, makes session IDs explicit tool parameters instead of auto-injecting them, and reifies language backends into a new LanguageBackendRegistry (Sep 18–19).

ChromeDevTools/chrome-devtools-mcp

Refactors call-type handling for stronger type safety (#2659, Sep 18), building on last week's chunked trace-buffer parser.

Unchanged in window: upstash/context7 (last release Sep 14), modelcontextprotocol/ext-skills (Sep 18 CODEOWNERS update already reported, no further changes), DeusData/codebase-memory-mcp (v0.11.0 already reported, no merges since Sep 16), github/github-mcp-server (reaction-removal tools already reported, no merges since Sep 16).

💡 Community — Workflows & Ideas

alibaba/open-code-review

Open-sources Alibaba's internal hybrid code-review tool — deterministic rule pipelines plus an LLM agent, OpenAI- and Anthropic-compatible — that reads a git diff and returns precise line-level comments; hits #1 on GitHub Trending Sep 16 with 3,200+ stars added that day, reaching 36k+ by Sep 18.

Tencent/BrowserSkill

Lets Claude Code, Cursor, Codex and other agents reuse an already-logged-in browser via a unified bsk CLI instead of separate API integrations; adds 1,300+ stars on Sep 18 trending.

📰 Quick Mentions

• Google opens Home to MCP agents — early access (Sep 16) lets Claude and other MCP clients read/control Google Home devices for Premium Advanced US subscribers; door unlocking and automations stay blocked.
• Anthropic's revenue pace tops $100B — annualized revenue run-rate reportedly passes $100B ahead of a possible November IPO (Sep 18).
• CCTV-affiliated account flags Anthropic privacy risks — alleges 13 privacy-policy revisions since 2023 increase cross-border data exposure to US agencies (Sep 19).
• anthropics/claude-plugins-official — adds incident.io and Wingspan partner plugins to the official marketplace (Sep 18).
• Piebald-AI/claude-code-system-prompts — tracks Claude Code's system prompt growing by roughly 5,234 tokens in v2.1.277 (Sep 18).