Claude Report — 2026-09-18

🚀 Anthropic Official

Claude Code v2.1.272–v2.1.276 (Sep 15–18)

Five releases ship in four days:

• v2.1.273 (Sep 15): adds LLM-gateway hint headers (x-claude-code-request-class, etc.), MCP-disconnect notifications, and forking of --remote-control sessions to run locally in the background; fixes a bypass where a subshell could hide a dangerous rm from permissions.blockReadsOutsideWorkingDirectories.
• v2.1.274 (Sep 17): adds a critical-memory-usage warning, CLAUDE_CODE_MCP_STARTUP_WAIT_MS, and self-healing for sessions stuck endlessly retrying "unexpected tool_use_id" errors.
• v2.1.275 (Sep 17): adds a send-now key (ctrl+enter) to interrupt and flush queued messages, syncs claude.ai account skills/plugins into terminal sessions (opt out via syncClaudeAiSkills/syncClaudeAiPlugins), adds /plugin install <plugin> --marketplace <source>, and hardens npm-sourced plugin installs (npm pack --ignore-scripts + integrity check) against malicious install scripts.
• v2.1.276 (Sep 18): hotfixes a v2.1.275 regression that broke every request behind a proxy/gateway with a 400 advisor_20260301 error.

claude-agent-sdk-typescript v0.3.272–v0.3.276 (Sep 15–18)

Syncs to each Claude Code release and adds a usage_report sibling on headless /usage results, mcpServer: {name, source} on canUseTool/tool-hook inputs, and fixes getSessionMessages()/forkSession() dropping or mis-keying messages sent while Claude was running a tool.

claude-agent-sdk-python v0.2.153–v0.2.156 (Sep 15–18)

Adds a snapshot field to SystemPromptPreset (plus a new SystemPromptCustom type) that freezes the system prompt at the first request for better prompt-cache hits across resumed sessions, then syncs its bundled CLI through v2.1.276.

Claude for Small Business expands to 43 workflows, 27 integrations (Sep 15)

Following 900,000 installs since its May launch, Anthropic adds workflows for lead follow-up, proposals and marketing campaigns (Shopify, Salesforce, Stripe, Zapier, and more) plus a fall schedule of training workshops and partner webinars.

Projects redesigned into a multi-thread Claude Code coordinator (Sep 17)

Claude now decomposes a single request across parallel cloud-session threads that share growing memory and a project file/artifact library, delegating tasks, reviewing outputs and assembling results automatically instead of requiring manual session splitting; beta starts with select Pro/Max Claude Code cloud-session users, expanding to Team/Enterprise and web/desktop later.

Life Sciences Verification Program (Sep 17)

Launches a beta, developed with the US government, that replaces real-time safety blocking with offline usage monitoring for vetted life-science organizations doing drug discovery, clinical development and manufacturing work; available now via the API console and Claude for Enterprise/Team, with individual Pro/Max access planned later.

Measurements for understanding the pace of AI development inside frontier labs (Sep 17)

Publishes three new transparency metrics (R&D share done by AI, oversight quality of AI agents, compute allocation) and discloses Claude now leads about 26% of Anthropic's own model R&D end-to-end under supervision, up from near zero at the start of the year.

How Claude is uplifting biomolecular modeling (Sep 17)

Claude accelerates 30+ open-source biology models roughly 4x on average with new FlashPairformer kernels (2.7–3.2x faster than prior standards), enables structure prediction beyond 10,000 tokens on a single GPU node, cuts a protein-design campaign's cost to about $150, and Anthropic open-sources the code and launches a $1M protein-design contest with Adaptyv Bio.

Compliance API adds Claude in Chrome session transcripts (Sep 18)

The Compliance API's local-session endpoints now return transcripts of Claude in Chrome sessions (product_surface: claude_in_chrome), in beta for Claude Enterprise orgs via the existing Compliance Access Key and read:compliance_user_data scope.

🔌 Claude Code Plugins

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

🛠️ Skills

eugeniughelbur/obsidian-second-brain

Ships a bootstrap style-flag feature (--style wiki|obsidian) and fixes a SessionStart hook silently skipping the vault manual, plus health-orphan detection and AppleDouble-file handling, across merges spanning Sep 15–18.

Unchanged in window: tt-a1i/archify (all commits predate Sep 17, already reported), ayghri/i-have-adhd (same), K-Dense-AI/scientific-agent-skills (no commits since Sep 11), affaan-m/ECC (Sep 17 activity is sponsor-doc housekeeping only, no functional change), agentskillexchange/skills (automated sync/regenerate bot commits only).

🤖 Agents & Subagents

None of the top items updated in window.

Unchanged in window: nicobailon/pi-subagents (last activity Sep 15, already reported), VoltAgent/awesome-claude-code-subagents, anthropics/commerce-agents.

🔗 MCPs & Integrations

oraios/serena

Ships a beta REPL feature, a persistent authentication secret for project-server communication, and restricted config-file permissions on POSIX, across roughly 30 commits Sep 15–17.

ChromeDevTools/chrome-devtools-mcp

Adds a chunked trace-buffer parser for large performance recordings (#2721, Sep 17), on top of sanitized telemetry and multi-file upload_file support already shipped.

upstash/context7

Documents native AWS Bedrock setup and clarifies regional Bedrock configuration and CLI login recovery, across three PRs merged Sep 16–17.

modelcontextprotocol/ext-skills

Updates CODEOWNERS for a new team assignment (Sep 18), continued housekeeping after the SEP-2640 Skills-over-MCP spec finalized Sep 16.

Unchanged in window: 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

cache-tax

Ships v2.1.0 and v2.1.1 (Sep 17) warning before a message triggers a costly prompt-cache cold rewrite — a cache write costs roughly 80x a read on Fable 5.1 — and keeps the 1-hour cache warm with periodic pings keyed by session id.

jev-router

Show HN (Sep 17): a loopback proxy that classifies Claude Code prompt complexity per turn and routes it to a fast/balanced/strong model tier while preserving the CLI's native tools, sessions and auth.

Claude Code from Source

Show HN (Sep 18, 16 points): an 18-chapter book reconstructing Claude Code's internal architecture — agent loop, tool execution, multi-agent orchestration, memory — as pseudocode, built from the CLI's npm source maps by 36 AI agents in about 6 hours.

📰 Quick Mentions

• Piebald-AI/claude-code-system-prompts — tracks Claude Code's full system prompt across releases; v2.1.274 added ~6,150 tokens (Sep 17), v2.1.275 trimmed ~1,237 tokens (Sep 17), and v2.1.276 made no prompt changes (Sep 18).
• anthropics/claude-plugins-official — adds Informatica for Claude Platform (Sep 16), then Leadfeeder and Chronograph GP/LP (Sep 17) to the official marketplace.
• Security researchers hack OpenAI using Claude — three researchers chain a libheif heap overflow with an SSO flaw using Claude Opus 5, reach OpenAI's internal GitHub and employee accounts in under 72 hours for under $3,000 in tokens, and responsibly disclose for a $6,500 bounty (Sep 18).