Claude Report — 2026-08-07

🚀 Anthropic Official

Claude Code v2.1.224

Ships Aug 7, 2026 (04:00 UTC), one day after the v2.1.223 security-patch release:

• Adds self-hosted environments (public beta, Team/Enterprise) — claude self-hosted-runner turns an org's own machines or containers into places where Claude Code web, mobile, and desktop sessions run, keeping code and build artifacts inside the org's network.
• Adds archive-based plugin installs: plugins can now install from a zip over HTTPS without git or npm, with optional SHA-256 pinning.
• Adds cross-session messaging (crossSessionInbound, dialogExpiry settings) so Claude Code sessions can message each other and discover peers via ListAgents (macOS/Linux).
• Adds sandbox credential masking: extract/decode: "jwt" with maskClaims, plus AWS SigV4 re-signing support.
• Adds ANTHROPIC_BEDROCK_REGION_PREFIX for cross-region Bedrock inference profiles; removes the 200-subagent spawn cap.
• Fixes long project-path session-directory collisions, sandbox filesystem denials, MCP tool bugs, and paste-content handling.

Agent SDKs ship matching releases

• claude-agent-sdk-python v0.2.132 (Aug 7): bundles CLI v2.1.224, no functional SDK changes.
• claude-agent-sdk-typescript v0.3.224 (Aug 7): adds the crossSessionInbound/dialogExpiry settings and a peer-send-message notification subkind; adds an archive plugin-source config variant (url + optional sha256); adds sandbox credential-masking fields (decode: 'jwt' with maskClaims, envVars.extract, awsPairs/sigv4); fixes long (>200 char) project paths resolving into another project's session directory, which had let /resume and session list/get/rename/tag/fork/delete cross project boundaries.

Improving Fable 5's biology safeguards

Published Aug 7, 2026: Anthropic rewrites the safety classifier's constitution gating biology queries, retrains it on updated data after soliciting internal and external expert feedback, and reports an approximately 85% reduction in biology-related false-positive blocks across product surfaces, while keeping blocks on dual-use virology, toxicology, and molecular-design content.

Tino Cuéllar joins Anthropic as Chief Global Affairs Officer

Published Aug 4, 2026: Mariano-Florentino "Tino" Cuéllar — former Carnegie Endowment president, California Supreme Court justice, and Stanford Freeman Spogli Institute director — becomes Anthropic's inaugural Chief Global Affairs Officer, overseeing policy, international strategic engagement, and government relationships.

🔌 Claude Code Plugins

Plugin install-rank cache unavailable this cycle — top-10 install-rank tracking skipped per this run's instructions.

🛠️ Skills

agentskillexchange/skills

Fixes wording in the security-scan section of the "OutageDeck dependency outage triage" skill and continues shipping the "shopify-app-review-triage-workflow" skill (Aug 6), alongside continuous bot-driven sync/regeneration commits running through Aug 7.

Unchanged in window: anthropics/skills (last Jul 24), Piebald-AI/claude-code-system-prompts (last Aug 6, no new capture since), JSONbored/awesome-claude (last Jul 31), alirezarezvani/claude-skills (last Jul 17).

🤖 Agents & Subagents

anthropics/claude-agent-sdk-typescript — v0.3.224

Ships (Aug 7) alongside the Claude Code CLI release with cross-session messaging settings, archive-based plugin config, and sandbox credential masking (full detail above).

nicobailon/pi-subagents — v0.42.1

Releases (Aug 6, 23:05 UTC) fixing subagent status-widget crashes in narrow terminals and async scripted-workflow timeouts, right after v0.42.0 (Aug 6, 18:39) split docs into focused guides and added native permission-system compatibility for child processes — both up from v0.41.0 (Aug 5).

AMAP-ML/LongHorizon-Harness

Merges v0.1.2 with unified computer-use plugin support (Aug 6) and v0.1.3 adding a final user-facing reply step run in the launch directory (Aug 7), continuing rapid iteration on its Manager/Executor/Auditor long-horizon harness.

Unchanged in window: Chachamaru127/claude-code-harness (last Aug 3), VoltAgent/awesome-claude-code-subagents (last Jul 31), wshobson/agents (last Jul 18), 0xmmo/crew (last Jul 22).

🔗 MCPs & Integrations

awslabs/mcp

Adds Anthropic MCP-directory-required annotations, icon, prompts, and support URL to redshift-mcp-server plus read-only tool annotations (Aug 7); migrates openapi-mcp-server to native FastMCP.from_openapi and fixes an aws-api-mcp-server security-policy issue via botocore.xform_name (Aug 6).

github/github-mcp-server

Adds a non-default find_duplicate tool gated by a duplicate_detection flag and support for removing issue types (Aug 7), after making search_issues semantic by default (Aug 6).

DeusData/codebase-memory-mcp

Merges a workspace-grants-and-manifest feature, a dry-run hook-refusal prediction fix, a Linux allocator-override fix, and a POSIX-spawn subprocess fix (Aug 7), building on Aug 6's per-project request manifest and mimalloc routing fixes.

Unchanged in window: oraios/serena (last Aug 5), modelcontextprotocol/python-sdk (last Jul 28, v2.0.0), modelcontextprotocol/typescript-sdk (last Jul 27), shlokkhemani/rabbithole (last Jul 29).

💡 Community — Workflows & Ideas

Anthropic confirms in-house AI chip design team

Confirms (Aug 5) it is assembling an internal silicon team — anchored by ex-OpenAI/Tesla-Dojo hardware lead Clive Chan — to co-design chips and Claude models together, targeting roughly 50% lower per-token inference costs, while continuing multi-vendor partnerships with Nvidia, AMD, Google, and Amazon.

Security research roundup: Claude Code and Gemini CLI flaws

Reports (Aug 7) a critical OS command-injection flaw in Gemini CLI (CVE-2026-12537, CVSS 10.0) via crafted .gemini/.env files, alongside a recap of Claude Code's already-patched data-exfiltration bug (CVE-2026-54316) that leaked API keys through Hugging Face's public download counter; both tools are patched, and users are urged to update and audit externally triggered workflows.

Keyv-linked npm worm plants malicious Claude Code hooks

Discloses (Aug 4) that a compromised Keyv npm package carries two payload paths via .claude/settings.json and .vscode/tasks.json, designed to trigger when developers open the repo in Claude Code or VS Code; both tools' default trust prompts block automatic execution, but the malicious files remained on the repo's main branch at time of reporting.

📰 Quick Mentions

• quemsah/awesome-claude-plugins — Automated n8n-workflow tracker (snapshot dated Aug 6, 2026) indexing 31,845 repositories for Claude Code plugin-adoption metrics.