Claude Report — 2026-09-05

🚀 Anthropic Official

Claude Code v2.1.259, v2.1.260, v2.1.261 (Sep 2–4)

v2.1.259 (Sep 2)

• Adds managedMcpServers so organizations can push HTTP/SSE MCP servers to every user, and --permission-prompts none for unattended headless hosts.
• Adds --json to claude plugin validate and GitLab merge-request recognition (MR !N) in tool summaries.
• Fixes concurrent sessions silently reverting each other's ~/.claude.json changes, and CLAUDE_CODE_MAX_CONTEXT_TOKENS being ignored for Vertex-style model IDs.
• Fixes managed settings going silently unenforced when the settings file can't be parsed — Claude Code now refuses to start and names the source.

v2.1.260 (Sep 3)

• Adds a fullscreen diff panel (/diff) showing live uncommitted changes, and a likely-cause hint for prompt-cache misses in /cost.
• Adds a text form of /advisor and /reload-plugins support for headless sessions.
• Fixes permission rules with parentheses being dropped as invalid, and zsh commands hiding command substitutions in REPORTTIME/REPORTMEMORY auto-approving instead of prompting.
• Fixes Fable 5.1 prompt caching not covering context attached after tool results, and SDK-provided MCP servers missing from initial turns.

v2.1.261 (Sep 4)

• Adds an "Organization policy" diagnostic line to /status/claude doctor, bashOutputMaxChars/taskOutputMaxChars (up to 128K), --append-subagent-system-prompt-file, and /skill-doctor to show unused skills and their context cost.
• Fixes the Bedrock setup wizard hanging indefinitely on unresponsive AWS credential helpers, and cloud sessions discarding a plugin synced from claude.ai when managed settings force-enable it.
• [VS Code] Adds an MCP server Add/Remove form directly in the IDE and a "Build a custom style" walkthrough.

ant apply — ant CLI v1.30.0 (Sep 3)

Ships a new command that creates and updates agents, environments, skills, memory stores, and deployments from files in a repo, committing a claude-lock.json lockfile so repeat runs update existing resources instead of duplicating them.

Claude Agent SDK TypeScript v0.3.259, v0.3.260, v0.3.261 (Sep 2–4)

Adds user_message_uuids and a permissionPrompts: 'none' option (v0.3.259), user_message_uuid on thinking-token messages plus four remote-session latency-breakdown fields (v0.3.260), and plugin delivery via stdin with a Symbol.dispose fix for Jest/Vitest (v0.3.261) — tracking Claude Code CLI parity through 2.1.261.

Claude Agent SDK Python v0.2.152 (Sep 2, unchanged since)

Bumps the bundled Claude CLI to 2.1.259 only; no release yet matches CLI 2.1.260/261, widening the gap with the TypeScript SDK.

🔌 Claude Code Plugins

Plugin tracking skipped this run — marketplace install-count cache unavailable (per mission note).

🛠️ Skills

tt-a1i/archify

Ships a CLI fix rejecting unknown options and extra positionals in render (#303), a preview fallback to polling after watcher errors (#275), and a canvas-containment fix for sequence segment labels (#299), across Sep 4–5.

Unchanged in window: anthropics/skills (frontend-design update still the last commit, Sep 3, already reported), agentskillexchange/skills (routine sync/regeneration commits only, plus a minor Tokenjuice fix), K-Dense-AI/scientific-agent-skills (last push Sep 2, already reported), alirezarezvani/claude-skills (no change since Aug 26).

🤖 Agents & Subagents

nicobailon/pi-subagents v0.65.1

Improves background-run reliability so sessions stay alive until work and result delivery finish, fixes Pi 0.85.0 compatibility, and preserves complete patches before worktree cleanup (Sep 4).

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — ships v0.3.261 (Sep 4) with stdin-based plugin delivery and a test-runner dispose fix.

Unchanged in window: claude-agent-sdk-python (still v0.2.152), VoltAgent/awesome-claude-code-subagents (README-only churn, no content change), AMAP-ML/LongHorizon-Harness (no push since Aug 20).

🔗 MCPs & Integrations

upstash/context7

Adds production HTTP controls (#3137), Docs7 attribution docs (#3138), and a version-packages release (#3102), building on Sep 2's VS Code/Devin/Copilot CLI setup targets and its Claude Code plugin authentication (Sep 4).

DeusData/codebase-memory-mcp

Adds an explicit quiet CLI output mode and makes MCP tool output lean and lossless, fixes logging defaults so the library stays at INFO while quiet becomes a process-level policy, and makes scoped file-search patterns portable (Sep 5).

oraios/serena

Fixes SolidLSP returning the wrong position from insert_text_at_position (#1842, Sep 4), following Sep 3's FreeBSD platform detection and language-server descendant reaping already reported.

Unchanged in window: github/github-mcp-server (same Sep 2–3 commits already reported: Agent Plugins 1.0 package, sanitizer/Markdown-fidelity fixes), modelcontextprotocol/typescript-sdk (same Sep 3 OAuth-indicator and Error.cause fixes already reported), modelcontextprotocol/servers (routine memory-server hardening commits Sep 2–3, no major new feature), punkpeye/awesome-mcp-servers (last commit Sep 1).

💡 Community — Workflows & Ideas

Portal by Spotify cuts Claude Code token usage 90%

Describes a "shunt" Claude Code plugin that hooks file reads and boilerplate generation to cheaper models (Gemini 2.5 Flash for bulk reads) via Spotify's Portal agent-runtime platform, keeping Claude's token ledger for reasoning-heavy work only.

14 MCP servers changed schemas within 27 hours, descriptions unchanged

Fingerprints 248 public MCP servers twice in a day and finds 17 tools altered their schema or annotations (e.g. a flipped readOnlyHint) with byte-identical descriptions — invisible to reviewers who only diff text, and no MCP client re-prompts on the change.

Claude's new system prompt bans reproducing song lyrics

Simon Willison flags a new clause refusing lyrics, poems, or book/article passages "in whole or in part," timed shortly after Sony Music Publishing and Warner Chappell's legal action against Anthropic; pre-1929 works remain exempt.

Context Engineering Kit — 13 Claude Code skill collections

Packages token-efficient Claude Code skills for reflexion loops, spec-driven development, subagent orchestration, and TDD anti-pattern detection, aimed at cutting hallucinations and improving consistency (surfaced on HN Sep 4).

Tool Definition Quality Score (TDQS) spec

Proposes a six-dimension, A–F grading framework for MCP tool definitions after finding 97% of tool descriptions have defects and well-described tools get selected roughly 2.6x more often (surfaced on HN Sep 3).

📰 Quick Mentions

• ChatGPT, Claude, and Grok all had outages at the same time (Sep 3) — a shared Azure East US failure took down Claude Mythos/Fable 5.1 and Opus 5 for roughly three hours.
• Anthropic has already raised $130B ahead of its IPO (Sep 4) — targets a late-September/early-October listing with Morgan Stanley, Goldman Sachs, and JPMorgan running the book.
• Show HN: Ccswitch (Sep 4) — a small CLI for switching between multiple Claude Code accounts.
• Hexis (Sep 4) — a tool for teams to collaborate on shared Claude Skills.
• Show HN: MCP Security Auditor (Sep 4) — a zero-execution static scanner for MCP server code.
• Sandboxing Claude CLI with Tart on Apple Silicon (Sep 4) — runs Claude Code inside lightweight Apple Silicon VMs for isolation.
• Claude Fable 5.1 and Mythos 5.1: The System Card (Sep 5) — a detailed community read of Anthropic's Sep 1 system card.
• Claude Code Release Pulse (Sep 4) — a third-party tracker summarizing Claude Code's release cadence.
• Show HN: Declick (Sep 4) — turns an OpenAPI spec, MCP server, or SQLite DB into a CLI.
• anthropics/skills trending (Sep 5) — gains 511 stars in a single day on GitHub's daily trending page.