Claude Report — 2026-08-20

🚀 Anthropic Official

Claude Code v2.1.236 (Aug 19) & v2.1.237 (Aug 20)

• v2.1.236 adds an ANTHROPIC_DEFAULT_MODEL env var to set the model new sessions start on, adds notify_when_idle for cross-session SendMessage (macOS/Linux), makes macOS sandbox wildcard read-deny rules take precedence and unbypassable by renaming denied files, and fixes clipboard copy, background sessions, local MCP logs after directory removal, and the fullscreen renderer's permanent-failure state (now falls back to the classic renderer).
• v2.1.237 fixes prompt caching for sessions using an LLM gateway or custom base URL, and adds a built-in "Concise" output style (results-first, no preamble/narration) selectable via /config → Output style.

Claude platform APIs reach GA (Aug 18-19)

• Files API is now GA (no beta header needed), adds file expiration, ids[] filtering, and pagination; storage capped at 1TB/org, 500 req/min.
• Agent Skills and the Skills API (/v1/skills) are now GA.
• Admin API user-management endpoints for Claude Enterprise (members, invites, groups, custom roles) are now GA.
• Claude Managed Agents gain allowed_domains/blocked_domains restrictions on web_search/web_fetch, plus max_content_tokens and user_location; self-hosted sandbox sessions can now attach persistent memory stores.
• Claude Console redesigns its session viewer (timeline minimap, per-request transcript grouping, Inspector panel) and renames Workbench to Playground.

Anthropic SDKs bump for the new GA features (Aug 19)

SDK Python ships v0.123.0, v0.124.0, v0.125.0; SDK TypeScript ships v0.118.0, v0.119.0, v0.120.0 — both add Files/Skills API GA support, memory-store and workspace-ID helpers, computer/browser-use toolsets, and managed-agent web-search domain config. Companion SDKs bump the same day: bedrock-sdk v0.33.0, vertex-sdk v0.19.5, google-cloud-sdk v0.0.10, foundry-sdk v0.4.4, aws-sdk v0.6.5.

Claude Agent SDK for Python v0.2.141 / v0.2.142 (Aug 19-20)

Bundles the updated Claude CLI (first v2.1.236, then v2.1.237) into the Agent SDK.

Claude Agent SDK for TypeScript v0.3.236 / v0.3.237 (Aug 19-20)

v0.3.236 lets PostToolUse hooks return hookSpecificOutput.classifierContext, read by the auto-mode permission classifier; v0.3.237 brings parity with Claude Code v2.1.237.

🔌 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 an "AShareHub" Chinese market-data skill (Aug 20), on top of the "SandBase" and "Atlas Cloud" skills already added Aug 19, plus continuous automated catalog syncs.

tt-a1i/archify

Fixes the viewer to preserve authored direction in Intent Trace (#90, Aug 19), on top of its Aug 17 v2.15.0 release.

Unchanged in window: Piebald-AI/claude-code-system-prompts (still the v2.1.235 system-prompt diff, Aug 19, already covered), anthropics/skills (still the academy-guide rename, Aug 18), alirezarezvani/claude-skills (no confirmed recent push).

🤖 Agents & Subagents

nicobailon/pi-subagents v0.52.0 / v0.52.1 (Aug 19-20)

Ships perf fixes to status-snapshot serialization, terminal-index paths, and workflow-fingerprint reuse, plus a subagent model-provider fallback fix.

anthropics/claude-agent-sdk-python (Aug 19-20)

Ships v0.2.141 then v0.2.142 — see Anthropic Official above.

anthropics/claude-agent-sdk-typescript (Aug 19-20)

Ships v0.3.236 then v0.3.237 — see Anthropic Official above.

AMAP-ML/LongHorizon-Harness

Merges three reliability fixes (Aug 20): guard-only failures now fail the round instead of the whole run (plus a new guard_exclude_paths option), recovers from isolated EBADF errors on the control bus, and fixes an AttributeError from extra_env not persisting on the agent; still on tagged release v0.1.6.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Aug 12).

🔗 MCPs & Integrations

github/github-mcp-server v1.10.0 / v1.10.1

Ships v1.10.0 (Aug 19) with a major security/reliability pass — confirmed repo deletion via form elicitation, bearer credentials restricted to configured GitHub authorities, enforced HTTPS for GHE hosts, fail-closed invalid --tools config, plus lockdown/rate-limit/cache-isolation hardening — then v1.10.1 (Aug 20) fixing an add_issue_comment schema-compatibility regression.

oraios/serena

Fixes a mid-indexing tsserver crash that was silently returning {} from find_referencing_symbols instead of surfacing the failure (PR #1848, Aug 20) — wires the existing window/logMessage crash signal into the indexing-wait logic so crashes now surface instead of getting masked.

modelcontextprotocol/python-sdk

Merges a commit flurry (Aug 17-18): fixes the HTTP 202 notification-acknowledgment path, gates external PRs in CI, and updates Client/StdioServerParameters and content-block typing.

Unchanged in window: modelcontextprotocol/typescript-sdk (still the Aug 18 auth token-refresh fix), DeusData/codebase-memory-mcp (still v0.10.8, Aug 19), shlokkhemani/rabbithole (still Aug 19 UI/CI work), awslabs/mcp (last release Aug 14).

💡 Community — Workflows & Ideas

Show HN: Do-over — undo for AI agent shell commands (Aug 20)

A Rust tool that hooks into Claude Code's hook system to snapshot files right before a destructive bash command (rm -rf, git reset, rsync) runs, protecting untracked/gitignored files and anything outside the repo.

Show HN: Pharos — a package manager for MCP servers (Aug 18)

A single Go-binary CLI that works like npm for MCP servers: searches across registries, installs with dependency resolution and lockfiles, audits installed servers against known vulnerabilities, and imports existing MCP client configs.

TokenBlast: tune Claude Code's environment flags (Aug 19)

A web UI with sliders and presets for cost, thinking, and context that generates a ~/.claude/settings.json, and diffs undocumented flags added/removed/restructured between Claude Code releases.

Show HN: Frugal Tokens — explore costs and usage across coding agents (Aug 19)

A dashboard analyzing per-session cost, cache-miss impact, overlapping sessions, and working-time estimates across coding-agent usage.

Show HN: Agent-codemode — scripts your agent writes can call your MCP servers (Aug 18)

Implements the "codemode" pattern, letting an agent write code that directly invokes MCP server functions instead of going through per-call tool-use round trips.

📰 Quick Mentions

• Feature request: support AGENTS.md in Claude Code (Aug 19, 267 pts) — GitHub issue debate over native support for the cross-vendor AGENTS.md convention vs. CLAUDE.md.
• Show HN: Runbook.v1 — governed, fail-closed workflow execution for MCP (Aug 18) — a spec/runtime for governed MCP workflow execution.
• Show HN: MCP server for Android drives apps via AI with no root (Aug 19) — controls Android apps through the accessibility tree with local PII redaction.
• Show HN: oc turns websites into micro CLIs for Claude Code (Aug 19) — wraps sites as compact CLIs so agents browse in hundreds of tokens instead of tens of thousands.
• Show HN: Flow — Claude Code pipeline from planning to merge (Aug 19) — chains triage, plan, worktree, code, verify, CI, review, and merge behind one approval gate.
• Show HN: Knownbase — an MCP server for persistent AI agent memory (Aug 19) — persistent cross-session memory for agents via MCP.
• Pragma — a Claude Code pipeline for iOS with 94 merged PRs (Aug 19) — an 8-stage spec-to-release slash-command pipeline for agent-driven iOS development.
• Show HN: Codewindow — picture-in-picture for terminal agents (Aug 18) — an Arc-style PiP widget that keeps a running coding-agent terminal visible across desktop switches.
• Show HN: Cronloop — run Claude Code or Codex on a schedule (Aug 18) — a scheduler service for cron-like coding-agent jobs.
• Sabrix-bench — MCP inspector and agent-loop latency profiler (Aug 19) — a Rust tool combining an MCP JSON-RPC inspector with a multi-turn agent latency benchmark.