Claude Report — 2026-09-26

🚀 Anthropic Official

Claude Marketplace launch (Sep 23)

Anthropic consolidates plugins/connectors (2,000+), Claude-powered agent products (CrowdStrike, Cursor, Harvey, Legora, Lovable, Snowflake), and service partners (Accenture, BCG, Deloitte) into a single discovery and procurement hub; customers can redirect committed Claude spend toward these listings.

Directory submission portal for plugins (Sep 25)

Opens plugin publishing to developers on paid plans: submissions (single MCP connector, or a GitHub-hosted bundle of MCP servers + skills) get automatic safety-scan validation, real-time review status, developer-controlled go-live, and post-launch install/discovery analytics. Adds support for the MCP 2.0 stateless-core spec, MCP Apps (interactive in-chat UI), and Enterprise Managed Auth (zero-touch OAuth).

Claude Code v2.1.283 (Sep 25)

Adds x-claude-code-prompt-id gateway hint headers for LLM-gateway request grouping, availableModelsMatch/deniedModels managed settings to block new or specific model releases, MCP/WebFetch/WebSearch tool output in OTel spans (OTEL_LOG_TOOL_CONTENT=1), and a /doctor prompt-audit command that audits CLAUDE.md, skills, agents, and commands for stale prompting patterns. Fixes SDK sessions losing deferred tool calls on early turn endings, MCP progress notifications dropped during background-task transitions, stateless remote MCP servers becoming unusable after a 404, and several plugin-management bugs (claude plugin validate, plugin details, marketplace remove).

claude-agent-sdk-typescript v0.3.283 (Sep 25)

Adds plugin_errors (with load-failure path) to the SDKSystemMessage type, fixes getSessionMessages()/forkSession() returning a rewound-away branch, surfaces turn warnings as system/informational stream-json messages, and changes set_max_thinking_tokens so omitting the field now preserves the current thinking budget (send null to reset). Ships same-day parity with Claude Code v2.1.283.

claude-agent-sdk-python v0.2.158 → v0.2.160 (Sep 23–25)

v0.2.158 adds a verbatim_prompts option that delivers user messages to the CLI exactly as written, skipping @path file expansion and slash-command dispatch to block untrusted-text injection. v0.2.160 fixes follow-up turns failing with "Stream closed" when a background subagent finished before its turn's result arrived, by keeping stdin open until the CLI reports idle (configurable via CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS).

Compliance API: Office Agents sessions out of beta (Sep 24)

Local session endpoints for Claude for Microsoft 365 (Excel, PowerPoint, Word, Outlook) graduate out of beta for Claude Enterprise organizations.

Compliance API: Activity Feed privacy update (Sep 24)

The Activity Feed no longer returns file names, project document names, or artifact titles — including retroactively for past activities; looking up a name by ID now requires a Compliance Access Key with the read:compliance_user_data scope.

Claude discovers a novel CRISPR-like enzyme system (Sep 23)

Roughly 950 Claude agents spent 21 hours and 210M tokens scanning a DNA-sequence database in Anthropic's biology lab; one agent flagged an uncharacterized "array-associated reverse transcriptase" system next to a CRISPR-like repeat array, later validated in wet-lab experiments.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

addyosmani/agent-skills

Bumps plugin manifests to 0.6.11 and merges a run of hardening PRs (Sep 25–26): guards spec/plan/todo artifact paths on every command surface, strips every negated trigger clause instead of just the first, enforces skill-anatomy layout rules, validates manifest versions against the root plugin.json, and scopes claude plugin eval via the manifest.

Unchanged in window: anthropics/skills (last commit Sep 24), cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills (last substantive commit Sep 13).

🤖 Agents & Subagents

None of the top 5 updated in window.

Unchanged in window: nicobailon/pi-subagents (last commit Sep 24), VoltAgent/awesome-claude-code-subagents (last commit Sep 21, README only), anthropics/commerce-agents (dormant since Aug 31).

🔗 MCPs & Integrations

upstash/context7

Documents final Search API pricing (Sep 25), continuing the Sep 23–24 Docs7/Search API documentation push.

DeusData/codebase-memory-mcp

Merges a Windows worker-recovery fix, a doclinks-markdown feature, and several smoke-test/daemon-bootstrap-sync fixes (Sep 25), on top of Sep 24's service-pattern-boundaries and directory-creation race-condition work.

ChromeDevTools/chrome-devtools-mcp

Fixes --blockedUrlPattern/--allowedUrlPattern rejecting hostname regexp groups and empty filePath being set as an empty string instead of undefined (Sep 25), its first commits since the Sep 23 v1.10.1 release.

Unchanged in window: oraios/serena (last commit Sep 24), github/github-mcp-server (last commit Sep 16), modelcontextprotocol/servers (last commit Sep 22).

💡 Community — Workflows & Ideas

Jevmem — automatic project memory for Claude Code

An open-source layer that auto-extracts decisions, constraints, bugs, and todos from Claude Code/Cursor/Codex sessions into a JEVMEM.md file and re-injects relevant lines next session; author claims ~300ms/$0.0001 per save-or-skip decision versus 2.8-4.3s/$0.005-0.013 for asking a model directly (Sep 26).

davila7/claude-code-templates

Adds three new mods: jev-auto-mode (a JSON-policy permission layer judged by an external classifier), agent-flow (a live tree view of subagents and the context passed between them), and git-sidebar (a lazygit-style worktree/branch pane) (Sep 25–26).

📰 Quick Mentions

• Recurse — Show HN — a framework for developing and deploying specialist agents faster (Sep 26).
• Claude Security Plugin v0.12.0 — version bump merged into anthropics/claude-plugins-official (Sep 25).
• code-modernization plugin — ~35 commits in one day covering rule consolidation, floating-point tolerance comparisons, and a live progress pane (Sep 25).