Claude Report — 2026-08-11
- An unreleased Claude research model raises the <a href="https://www.anthropic.com/research/riemann-zeta">Riemann zeta zero lower bound from 41.6% to 67.2%</a>, using a ~60-subagent swarm over 31M output tokens.
- Claude ships mandatory <a href="https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content">content watermarking</a> across API, Claude, Claude Code, and Cowork to comply with the EU AI Act.
- Claude Code hits <a href="https://code.claude.com/docs/en/changelog">v2.1.227</a>; the Agent SDK and <a href="https://github.com/nicobailon/pi-subagents/releases/tag/v0.46.0">pi-subagents v0.46.0</a> ship parity/feature releases in step.
- MCP ecosystem churns: <a href="https://github.com/awslabs/mcp/releases">awslabs/mcp</a> patches a critical websocket CVE, <a href="https://github.com/github/github-mcp-server/releases/tag/v1.9.0">github-mcp-server hits v1.9.0</a>, DeusData keeps hardening its memory MCP.
- Community flags Claude Code's up-to-<a href="https://quesma.com/blog/claude-code-pricing-for-enterprise/">40x enterprise token pricing</a> vs. seat plans; <a href="https://github.com/activeing123/mcptoon">mcptoon</a> claims 97% token savings on MCP tool discovery.
🚀 Anthropic Official
Learning More About Claude's Mathematical Capabilities
An unreleased Claude research model, run as a ~60-subagent swarm inside Claude Code over a day and a half and 31M output tokens, raised the proven lower bound for the fraction of Riemann zeta zeros satisfying the Riemann hypothesis from 41.6% to 67.2% — the largest single-step advance on that bound in its history. It is not a proof of the hypothesis itself and hasn't cleared conventional peer review, but the result was formally verified in Lean and reviewed by external number theorists Brian Conrey and Dan Goldston; the Lean proof is public.
Claude Adds Content Watermarking for EU AI Act Compliance
Anthropic starts embedding imperceptible text watermarks and C2PA-signed provenance metadata (for .svg/.png/.jpg) in Claude-generated content, applying across the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, plus AWS/Google Cloud partners. Models launched Aug 2, 2026 or later support marking at launch; older models are transitioning, and a detected mark only indicates content "may have been processed by Claude," not full provenance.
Claude Code v2.1.226 and v2.1.227
- v2.1.226 (Aug 8): general bug fixes and reliability improvements.
- v2.1.227 (Aug 10): fixes feature flags evaluating without subscription tier on expired-token session starts (wrongly nudging Max users toward Fable usage credits), fixes Bash commands failing under
claude-code-actionwithallowed_non_write_userson GitHub-hosted runners, fixes/tuirestoring a rewound conversation, and improves the slash-command menu and event-loop performance.
Claude Code Adds Cross-Session Messaging (macOS/Linux)
Separate Claude Code sessions can now message each other directly — sending a task summary (not full history/files) so one session can warn another about a breaking change or answer a blocking question. Same-machine messages travel over per-session Unix domain sockets and never reach Anthropic's servers; inbound messages can't approve permission prompts or run slash commands. Unavailable on Bedrock, GCP Agent Platform, AWS Claude Platform, and Microsoft Foundry at launch.
🔌 Claude Code Plugins
Plugin install-rank cache unavailable this cycle — top-10 install-rank tracking skipped per this run's instructions.
🛠️ Skills
agentskillexchange/skills
Adds three new named skills in the window — an ImagineVid AI generation skill (Aug 8), a Unified AI System gateway skill (Aug 10), and a BRAINIALL diarized-transcription skill (Aug 11) — alongside its usual multiple-times-daily automated sync.
Piebald-AI/claude-code-system-prompts
Mirrors Claude Code's v2.1.226 (Aug 8) and v2.1.227 (Aug 10) system-prompt/changelog content, keeping its extracted-prompts archive in sync with the two latest CLI releases.
Unchanged in window: anthropics/skills (last Aug 7, already reported), JSONbored/awesome-claude (last Jul 31), alirezarezvani/claude-skills (last Jul 17).
🤖 Agents & Subagents
nicobailon/pi-subagents
Ships v0.45.1 and v0.45.2 (Aug 9-10, already noted yesterday) plus a new v0.46.0 (Aug 11) adding prompts.render(ref, vars?) for scoped prompt fragments, a versioned TypeScript API for project panes, and improved workflow child-visibility/status reporting.
anthropics/claude-agent-sdk-typescript
Ships v0.3.227 (Aug 10) to bring the SDK to parity with Claude Code v2.1.227.
Unchanged in window: AMAP-ML/LongHorizon-Harness (last Aug 10, already reported), Chachamaru127/claude-code-harness (last Aug 8, already reported), VoltAgent/awesome-claude-code-subagents (last Jul 31), wshobson/agents (last Jul 18), 0xmmo/crew (last Jul 22).
🔗 MCPs & Integrations
DeusData/codebase-memory-mcp
Keeps shipping through the window: makes every MCP tool reply client-usable via structuredContent and pipelining (Aug 10), fixes Windows RSS diagnostics and SQLite path normalization (Aug 10), and merges idempotent publish-registries handling plus a twine/Metadata-Version 2.5 fix (Aug 11).
awslabs/mcp
Ships release 2026.08.20260810225031 (Aug 11, 00:09 UTC) migrating openapi-mcp-server to FastMCP's native OpenAPI support, adding billing-credits tooling, and patching the critical websocket-driver CVE-2026-54466 alongside other dependency security fixes.
github/github-mcp-server
Ships v1.9.0 (Aug 10), bundling the descending list_label ordering with issue-type removal support, a gated find_duplicate tool, semantic-by-default search_issues, issue_read returning closing PRs, and a docs clarification that create_or_update_file content is plain text, not base64.
Unchanged in window: oraios/serena (last Aug 9 v1.7.0, already reported), shlokkhemani/rabbithole (last Aug 9, already reported), modelcontextprotocol/python-sdk (last Jul 28, v2.0.0), modelcontextprotocol/typescript-sdk (last Jul 27).
💡 Community — Workflows & Ideas
Show HN: mcptoon — a token-efficient MCP CLI client
Converts MCP JSON tool output to a compact TOON format, claiming roughly 97% fewer tokens on tool discovery and 40-60% fewer on results — the top-scoring MCP-tagged Show HN of the window (48 points, Aug 11).
Claude Code pricing: same tokens, same model, up to 40x the price
Argues Anthropic subsidizes seat-based Max/Team plans as a market-segmentation loss-leader while metered API/Enterprise billing exposes the true cost of agentic coding, with organizations switching from seats to tokens reporting bills 3x to 40x higher for equivalent work.
Claude could watermark text without invisible characters
Speculates that Claude's new text watermark works by subtly biasing next-token selection toward a secret pattern rather than hiding special characters, making the signal survive copy-paste and minor edits because it lives in word choice itself.
Show HN: Airship — a Figma-like visual editor for Claude Code, Codex, and OpenCode
Connects to a running dev server so a user can click an on-screen element, describe a change in natural language, and have Claude Code, Codex, or OpenCode edit the underlying source directly.
What changes Claude Code behaviour in CLAUDE.md (with a free generator)
Ships a generator for CLAUDE.md files that captures project-specific, non-obvious conventions a newcomer "couldn't work out in two minutes of reading," aimed at making Claude Code's automatic per-run context more effective.
📰 Quick Mentions
- Show HN: AI Pulse — a fake LED strip beside the macOS Dock that shows live coding-agent status (15 points, Aug 10).
- Show HN: Keen Code — an "agentic-engineered" coding agent positioned against Claude Code and Codex (Aug 10).
- GitHub issue: Claude ignores instructions to stop adding verbose code comments — an open Claude Code bug report gaining HN traction (6 points, Aug 11).
- Builder Radar weekly — the MCP SDK logs 52.1M weekly npm downloads vs. 29.8M for Anthropic's own SDK, for the week of Aug 9.
- Show HN: allmcps-server — an MCP server that itself serves as a directory/gateway to other MCP servers (Aug 10).
- Show HN: Graph2agent — turns Mermaid diagrams into agent-readable explanations (5 points, Aug 10).
- Show HN: Ungate — uses Claude and GPT subscriptions inside Cursor without incurring separate API costs (Aug 11).
- Show HN: TRMNL X e-ink usage screen — displays Claude Code and Codex usage stats on an e-ink device (Aug 11).
- pi-gpt-search — reverse-engineers Codex's web search so it works with Claude and local models too (Aug 11).
- MCP for live NBA game recaps — an MCP server pairing box scores with Reddit game-thread commentary (Aug 10).