Claude Report — 2026-08-22
- Claude Code hits v2.1.239: US-inference cost premium shown in estimates, `/claude-api upgrade`, ~40 fixes (Aug 21).
- Files, Agent Skills, and Admin user-management APIs exit beta; Managed Agents gain domain-restricted web tools (Aug 19).
- Claude Agent SDK TypeScript v0.3.239 adds cost multiplier and background-subagent cost reporting (Aug 21).
- Anthropic launches Claude Security for defenders, pairing Mythos 5 with a $35M open-source patching fund (Aug 21).
- pi-subagents ships v0.53.0 and v0.54.0; a cross-model "debuzz" skill tops Hacker News (Aug 20-21).
🚀 Anthropic Official
Claude Code v2.1.239 (Aug 21)
- Cost estimates now include the 1.1× US-only-inference premium for data-residency workspaces, and shows a one-time fullscreen-renderer offer on Bedrock/Vertex/Foundry setups.
- Adds
/claude-api upgradeto migrate Python projects to theanthropic1.x SDK; cloud-synced plugins now display asname@synced. - Fixes Alpine/musl clipboard, paste, and audio add-ons, plus roughly 40 other bugs including Bedrock proxy double-billing,
/resumepath errors, and Windows cross-sessionSendMessage/ListAgentsreliability.
Claude Agent SDK TypeScript v0.3.239 (Aug 21)
Adds the 1.1× US-only-inference multiplier to total_cost_usd/modelUsage.costUSD, makes background subagents report cost/duration/usage at release time instead of turn-end, fixes SYSTEM_PROMPT_DYNAMIC_BOUNDARY leaking as literal text on Bedrock/Vertex/Foundry/gateway providers, and adds a background_tasks_changed snapshot on host reconnect.
Claude Security launches with Claude Mythos 5 (Aug 21)
Anthropic ships Claude Security, an enterprise product that scans codebases and proposes patches with CWE categorization, confidence ratings, and severity scoring; pairs it with partner-embedded Mythos 5 deployment, a new Cyber Verification Program, and a $35M "Defender Advantage Fund" for open-source security patching.
Files API reaches GA (Aug 19)
The files-api-2025-04-14 beta header is no longer required; adds file expiration and page/ids[] pagination for listing files.
Agent Skills API reaches GA (Aug 19)
The skills-2025-10-02 beta header is no longer required to use Agent Skills.
Admin API user management reaches GA for Claude Enterprise (Aug 19)
The ce-user-management-2026-07-13 beta header is no longer required for group and custom-role requests.
Managed Agents add domain-restricted web tools (Aug 19)
Adds allowed_domains/blocked_domains configuration to web_search/web_fetch on agent_toolset_20260401.
Claude Console session viewer redesigned (Aug 19)
Adds a timeline minimap, per-model-request transcript grouping, and a new Inspector panel for debugging agent runs.
Anthropic SDK Python v0.123.0–v0.125.0 (Aug 19)
Brings Files/Skills APIs to GA in the SDK, adds computer-use and browser-use toolsets, managed-agents web search config, and self-hosted sandbox memory support — the release train leading into v1.0.0.
Anthropic SDK TypeScript v0.119.0–v0.120.0 (Aug 19)
Ships TypeScript parity with the Python SDK's GA Files/Skills APIs, computer-use/browser-use toolsets, and managed-agents/sandbox-memory features, alongside matching Bedrock, Vertex, Foundry, AWS, and Google Cloud sub-package bumps.
🔌 Claude Code Plugins
Plugin install-rank cache unavailable this cycle — top-10 install-rank tracking skipped per this run's instructions.
🛠️ Skills
tt-a1i/archify
Fixes Semantic Radar collision/dragging handling and route/boundary auto-rendering, and adds a RAG pipeline architecture example to its authoring cookbook (Aug 19-21).
Piebald-AI/claude-code-system-prompts
Publishes system-prompt token-footprint diffs for Claude Code v2.1.235 through v2.1.239 as each version ships (Aug 19-21).
agentskillexchange/skills
Adds several new marketplace skills — including a MuAPI image-generation skill, an AShareHub China-market data skill, and a SandBase multi-source research skill — via near-continuous automated sync through Aug 22.
anthropics/skills
Updates the claude-api skill with a Python SDK 0.x-to-1.x upgrade guide (Aug 21).
alirezarezvani/claude-skills
Merges a dev-branch update into main (PR #970, Aug 21).
🤖 Agents & Subagents
nicobailon/pi-subagents v0.53.0 → v0.54.0 (Aug 20-21)
Adds per-agent model scopes, an RPC schedule-management API, a TTL-based model-exclusion store for fallback candidates, and a runtime agent-registration API.
anthropics/claude-agent-sdk-typescript (Aug 21)
Ships v0.3.239 — see Anthropic Official above.
AMAP-ML/LongHorizon-Harness v0.1.7 (Aug 20)
Lets the dashboard continue a finished run as a conversation, fixes guard-only failures wrongly failing whole runs, and fixes isolated EBADF crashes in control-bus reads.
Unchanged in window: anthropics/claude-agent-sdk-python (still v0.2.143, Aug 20, already covered), VoltAgent/awesome-claude-code-subagents (last commit Aug 12).
🔗 MCPs & Integrations
oraios/serena
Fixes a tsserver crash that produced silent empty results (PR #1848, Aug 20).
DeusData/codebase-memory-mcp
Adds dbt/Jinja lineage extraction, Swift protocol-method indexing, and first-class SQL Table/View lineage nodes on top of the v0.10.8 release (Aug 20-21).
Unchanged in window: modelcontextprotocol/python-sdk (same SSE/OAuth body-size-limit and TypedDict fixes already covered), github/github-mcp-server (still v1.10.1, Aug 20, already covered), modelcontextprotocol/typescript-sdk (last commit Aug 18). Also tracked as alternates: shlokkhemani/rabbithole (last activity Aug 19), awslabs/mcp (commits Aug 20-21 but no new release tag).
💡 Community — Workflows & Ideas
Claudette (nobuzz) — make Claude stop talking like a BuzzFeed article (Aug 21)
A Claude Code skill pipes Claude's response to a different model (Gemini) for a plain-English rewrite in one of three audience modes, then prints that output verbatim without letting Claude re-polish it — a reusable cross-model critique pattern.
Frigade replaces its Grafana stack with a Claude Code skill (Aug 21)
Retires 24 hand-built dashboards for a skill that answers analytics questions over a read-only Postgres replica and posts scheduled Slack summaries, trading dashboard maintenance for occasional wrong answers.
OzBrain — a shared MCP knowledge store across agents (Aug 21)
Lets Claude, ChatGPT, Cursor, and Claude Code read/write the same structured knowledge base via one MCP connector, with row-level security and conflict detection when agents write contradictory information.
Proliferate — self-hostable multi-agent coding IDE (Aug 21)
Orchestrates Claude Code, Codex, OpenCode, Cursor, and Grok in parallel, giving each task an isolated git worktree, branch, and terminal.
Claude Code Hermit — 24/7 autonomous agent daemon (Aug 21)
Wraps Claude Code in a persistent daemon that only polls on filesystem changes (near-zero idle cost), with Discord/Telegram remote control and a native Artifact dashboard.
📰 Quick Mentions
- Vesta (Aug 21) — maintains a live code graph alongside Claude Code and reports a 38.8% cost reduction from proactive structural context injection.
- Desktop Vibe Fly (Aug 21) — a real neuroscience fly-brain simulation that treats AGENTS.md/CLAUDE.md files on disk as "odor sources."