New Tooling & Ecosystem (Q1-Q2 2026)

Pillar: new-tooling-2026 | Date: April 2026
Scope: New MCPs released since Jan 2026 worth evaluating: design system MCPs, accessibility MCPs (axe-core MCP, Pa11y), observability MCPs, vector store MCPs, agentic testing MCPs, browser automation beyond Playwright. New skills/commands ecosystems (community plugins, Anthropic official skill registry). New hook patterns published (PreToolUse/PostToolUse/Stop/SessionStart innovations). Anthropic harness team roadmap signals for next 1-2 quarters. What Jacob should NOT reinvent because it is shipping.
Sources: 63 gathered, consolidated, synthesized.

Table of Contents

1. Accessibility MCPs
2. Design System MCPs
3. Observability MCPs & Native OpenTelemetry
4. Vector Store MCPs
5. Browser Automation Beyond Playwright
6. Plugin & Skills Ecosystem
7. Hook System Innovations (Q1–Q2 2026)
8. Claude Code Platform Releases Q1–Q2 2026
9. Anthropic Platform Architecture & Roadmap
10. MCP Ecosystem Scale & Directory
11. What Not to Reinvent
12. Gaps & Missing Coverage

Section 1: Accessibility MCPs

Three distinct accessibility MCP solutions are now available for Claude Code: Deque's enterprise Axe MCP Server (GA February 11, 2026), a community open-source a11y-mcp (40 stars), and the Community-Access/accessibility-agents bundle (24 tools, 11 specialist agents).^[3][36][54] All axe-core-based tools share the same fundamental ceiling: automated scanning detects only 30–40% of accessibility issues; manual audits remain required for complete WCAG coverage.^[25]

Deque Axe MCP Server

Launched at Axe-con 2025, GA February 11, 2026, included at no additional cost in Axe DevTools for Web licenses.^[3][35] Available at https://github.com/dequelabs/axe-mcp-server-public. Explicitly confirmed for: GitHub Copilot, Cursor, Claude Code, VS Code.^[3]

Core workflow: analyze → remediate → validate, with configurable auto-validation of fixes. Connects to Deque University's WCAG-aligned knowledge base. CTO Dylan Barrell: "Developers can contribute to accessibility earlier in the software development lifecycle, all while using existing tools."^[3]

Roadmap: expanding to power AI agents across the full dev lifecycle; automated intelligent guided tests; advanced rules for increased issue detection.^[3]

Key finding: The Axe MCP Server is the only accessibility tool with confirmed Claude Code support and enterprise backing — but its 30–40% automated detection ceiling means it is a triage accelerator, not a compliance replacement.^[25]

Accessibility Tool Comparison

Pa11y Status

No dedicated Pa11y MCP server exists in 2026. Pa11y remains a CLI tool for batch URL testing via pa11y-ci, best integrated as a GitHub Actions step or hook script.^[25] The community a11y-mcp uses axe-core, not Pa11y. MCP skills on mcpmarket.com covering accessibility (e.g., "Accessibility Testing Skill for Claude Code | axe-core") also use axe-core exclusively.^[25]

Section 2: Design System MCPs

The design system MCP space consolidated around three production-grade offerings in Q1–Q2 2026: Figma's official MCP server (production since June 2025, expanded February 17, 2026), a Figma-native "Create Design System Rules" skill, and Storybook's MCP server (v10.3+, React only). Together they enable a full code→Figma→code roundtrip.^[4][27][38]

Figma MCP Server — Full Roundtrip (February 2026)

Figma MCP Server launched in Dev Mode in June 2025. On February 17, 2026, a new code-to-Figma direction was added, completing the full roundtrip.^[4][37][55] Claude Code is listed as an official Figma MCP partner.

Full roundtrip: Claude Code generates code → code runs in browser → screenshot captured → Figma frame created → designers iterate → Figma MCP feeds updated specs back to Claude Code.^[4]

Code Connect: Links existing codebase components to Figma designs, enabling side-by-side comparisons, annotation, and variant exploration without re-coding.^[55]

Community extension: southleft/figma-console-mcp — "Your design system as an API. Connect AI to Figma for extraction, creation, and debugging."^[4]

Key finding: The Figma MCP's February 2026 code-to-Figma direction completes a bidirectional loop — development artifacts can now flow into design tools without manual redraw, eliminating the traditional dev→designer handoff bottleneck.^[4]

Figma "Create Design System Rules" Skill

Official Figma Developer Docs skill generating custom design system rules that guide AI coding agents to produce consistent code when implementing Figma designs.^[38] Output targets: CLAUDE.md for Claude Code, .cursor/rules/figma-design-system.mdc for Cursor, AGENTS.md for Codex CLI.^[38]

5-step workflow: (1) Run tool with language/framework parameters → (2) Analyze codebase for component organization, styling, patterns → (3) Generate rules covering components, styling, Figma integration, assets, conventions → (4) Save to agent config → (5) Validate and iterate with test implementations.^[38]

Critical rules use "IMPORTANT:" prefix for conventions that must never be violated (e.g., "Never hardcode colors — use tokens from [location]").^[38]

Storybook MCP Server (v10.3+)

Storybook's official MCP server ships at http://localhost:6006/mcp in v10.3+. Connects AI agents to Storybook instances to understand components, generate stories, run tests, validate UI consistency.^[27] Current limitation: React projects only — Vue, Angular, Web Components, Svelte support planned.^[27]

Dual value: Storybook MCP is simultaneously a design system MCP (component discovery, documentation) and an agentic testing MCP (run-story-tests + accessibility checks).^[27]

See also: Agentic Testing pillar for Storybook's run-story-tests integration in CI pipelines

Section 3: Observability MCPs & Native OpenTelemetry

Claude Code now ships with built-in OTLP telemetry — no external wrapper required. Five additional observability MCPs are available for querying external systems: Datadog (preview), Grafana (v0.12.0, April 23, 2026), Traceloop OTel MCP (v0.2.2, 185 stars), community claude_telemetry wrapper, and New Relic's agentic platform (February 24, 2026).^{[29][19][28][39]}

Claude Code Native OpenTelemetry

Enable with CLAUDE_CODE_ENABLE_TELEMETRY=1. Three independent signal types, all OTLP-standard.^[29] Improved in v2.1.117 (April 22, 2026) with user_prompt and command_name events added to traces.^[9]

Trace Span Hierarchy

Privacy defaults: Prompt and content NOT recorded by default. Opt-in via: OTEL_LOG_USER_PROMPTS, OTEL_LOG_TOOL_DETAILS, OTEL_LOG_TOOL_CONTENT (truncated at 60KB).^[29] Subagent spans nest under parent claude_code.tool — full delegation chain visible in one trace. W3C trace context propagated into CLI subprocess and forwarded to Bash commands.^[29]

Supported backends: Honeycomb, Datadog, Grafana, Langfuse, self-hosted OTLP collector.^[29]

Key finding: Claude Code's native OTLP removes any need to build a separate observability wrapper — the instrumentation ships in the binary. The remaining gap is on the query/alert side, which Datadog MCP and Grafana MCP address.^[29][5][19]

Datadog MCP Server

Announced at Datadog DASH conference (June 2025), now in preview.^[39] Partners: Claude Code (confirmed), Cursor, OpenAI Codex CLI (partnership announced), Block's Goose.^[5][39] Supports HIPAA-compliant environments with user-based RBAC.^[5]

Grafana MCP Server

Official Grafana tool at grafana/mcp-grafana. 2,900 stars, 346 forks. v0.12.0 released April 23, 2026. Go (94.3%), Apache 2.0.^[19]

Tools by category: Dashboard operations (search, retrieve, patch, extract panel queries via JSONPath); Prometheus (PromQL instant/range queries, metric metadata, label discovery, histogram percentiles); Loki (LogQL queries, log patterns, label metadata); Alert rules; Grafana Incident creation/tracking; Sift investigations (error patterns, slow requests); OnCall (schedule/shift management); Rendering (Dashboard/panel PNG export); Deeplink generation.^[19]

Context-aware design: get dashboard summary and get dashboard property (JSONPath) tools minimize token cost vs. full dashboard JSON retrieval.^[19]

Traceloop OpenTelemetry MCP Server

GitHub: traceloop/opentelemetry-mcp-server. Version 0.2.2 (February 8, 2026). 185 stars, 15 forks. Apache 2.0. Unified MCP server for querying OTel traces across Jaeger, Grafana Tempo, and Traceloop Cloud.^[28]

LLM-specialized tools with OpenLLMetry semantic convention support:^[28] get_llm_usage (aggregate token usage across services/models), get_llm_model_stats (compare model performance), get_llm_expensive_traces (highest token-usage operations), get_llm_slow_traces (performance bottlenecks). Complementary to Datadog/Grafana for pure OTel trace querying.^[28]

Community & Emerging Observability

claude_telemetry (TechNickAI/claude_telemetry): Drop-in replacement swapping claude command for claudia. Logs tool calls, token usage, costs, execution traces. Backends: Logfire, Sentry, Honeycomb, Datadog. Alternative to full OTLP stack configuration.^[56]

New Relic Agentic Platform (February 24, 2026): No-code agentic platform for data observability agents. MCP support for connecting AI applications to data sources. Fleet management for OTel collectors. Focused on observability-persona agents, not general-purpose.^[5]

Section 4: Vector Store MCPs

Four vector store MCPs are in active use in 2026: Zilliz claude-context (9,900+ stars — dominant for semantic code search), Qdrant official MCP (1,400 stars — semantic memory layer), Pinecone Plugin (official, February 11, 2026 — RAG + assistant), and bobmatnyc/mcp-vector-search (ChromaDB-backed CLI-first search).^{[42][6][30][57]}

Zilliz Claude Context — Semantic Code Search

GitHub: zilliztech/claude-context. 9,900+ stars, 751 forks — highest adoption of any semantic code MCP.^[42] Benchmarked ~40% token reduction vs. traditional context loading. Handles millions of lines of code (enterprise-scale). Supports 15+ AI coding tools with Claude Code as primary.^[42]

Architecture: Hybrid search (BM25 keyword + dense vector embeddings), AST-based code chunking for 8+ languages (TypeScript, Python, Java, C++, C#, Go, Rust, etc.), incremental indexing via Merkle trees.^[42] Embedding providers: OpenAI (text-embedding-3-small/large), VoyageAI (voyage-code-3), Ollama (local), Google Gemini.^[42]

Setup: claude mcp add claude-context — requires free Zilliz Cloud account + embedding API key.^[42]

Community fork: danielbowne/claude-context — local LanceDB backend, no cloud required.^[57]

Contextual chunking pattern (from Anthropic research, applied by both Zilliz and bobmatnyc): Prepending compact metadata header to each chunk before embedding reduces retrieval failures by 35–49%.^[57]

Key finding: Modern coding agents (Claude Code, Cursor, Devin) are NOT primarily using vector RAG for code retrieval — they use file system navigation, grep, and AST parsing. Vector stores shine for knowledge bases and documentation where explicit structural relationships are absent. The 9,900-star adoption of claude-context suggests vector search adds value primarily when codebases exceed context window capacity.^[6][42]

Vector Store MCP Comparison

Qdrant Official MCP

qdrant/mcp-server-qdrant, v0.8.1 (December 2025), 1,400 stars, 268 forks, Apache 2.0.^[6] Two core tools: qdrant-store (store information with metadata to collections) and qdrant-find (semantic similarity search, up to 10 results by default).^[6] Default embedding: FastEmbed with sentence-transformers/all-MiniLM-L6-v2. Transport options: stdio, SSE, streamable HTTP.^[41] Positioned as semantic MEMORY layer — distinct from code navigation use case.^[6]

Pinecone Plugin (February 11, 2026)

Published via Anthropic's Claude Code Plugin Marketplace, February 11, 2026.^[30] Install: claude plugin install pinecone. Slash commands: /pinecone:query, /pinecone:assistant-create, /pinecone:assistant-upload, /pinecone:assistant-sync, /pinecone:assistant-chat.^[30] Pinecone Assistant supports upload of PDF, Markdown, TXT, DOCX, JSON files with cited answers and page references. Alternative MCP approach via npx @pinecone-database/mcp with cascading-search (multi-index with deduplication and reranking).^[30]

Section 5: Browser Automation Beyond Playwright

Five browser automation MCPs are production-ready in 2026, with Agent-Browser (Vercel Labs, 29,500+ stars) and Chrome DevTools MCP (Google official, 37,400 stars) as the major 2026 additions. Token efficiency ranges from 114K tokens/task (Playwright MCP) to 7,800 tokens/task (Agent-Browser) — a 14.7× spread that determines which tool is viable in constrained contexts.^[10][59]

Key finding: Industry consensus for 2026: "Start with Playwright MCP as default; choose Browserbase for cloud scale, mcp-chrome for working inside an already-logged-in browser, Browser Use for persistent agent sessions, and Chrome DevTools MCP for debugging and performance audits."^[31]

Token Efficiency Comparison

Chrome DevTools MCP (Google Official — April 2026)

Maintained by Google's ChromeDevTools team. 37,400 stars, 2,300 forks. v0.23.0, April 22, 2026. TypeScript (95.3%), Node.js 20.19+.^[10]

34 tools across 7 categories: Input automation (9 tools), Navigation (6), Emulation (2), Performance (3 — trace recording and analysis), Network (2), Debugging (6 — script execution, console, Lighthouse audits, screenshots), Extensions (5), Memory (1).^[10]

Installation: Plugin (MCP + Skills bundled) or MCP-only (claude mcp add chrome-devtools). "Slim mode" with just 3 tools for basic tasks.^[10]

Positioning: Specialized for QA, performance audits, debugging — NOT general automation. Best when you need Lighthouse, network inspection, or performance tracing.^[10]

Agent-Browser (Vercel Labs)

29,500+ GitHub stars — most-starred browser tool in 2026.^[10] Written in Rust for speed and minimal overhead. Innovation: generates accessibility tree, assigns @refs to interactive elements.^[47]

• 5.7× more tests in same context budget vs Playwright MCP^[47]
• 14.7× token reduction vs Playwright MCP (~7,800 vs ~114,000 tokens)^[59]
• Native CLI giving any AI model composable browser control via CDP^[59]

"Ralph Wiggum Loop" Pattern (Vercel/Pulumi, 2026): Self-verifying agent — make change → take snapshot → evaluate vs expected state → if wrong, correct and loop. No human in loop.^[59]

Stagehand v3 (Browserbase, February 2026)

Complete rewrite with AI-native architecture. Talks directly to browser via Chrome DevTools Protocol (CDP). 44% faster than v2. Multi-language support, driver-agnostic (not tied to Playwright). Three primitives: act (perform action), extract (pull data), observe (understand screen).^[10]

Browser Automation MCP Landscape

Shadow DOM limitation (2026): Modern design systems (Shoelace, Lit) hide elements in shadow roots. Accessibility tree snapshots — used by Playwright MCP, Agent-Browser, and Chrome DevTools MCP — cannot see them.^[47]

Section 6: Plugin & Skills Ecosystem

The Claude Code plugin ecosystem reached 4,200+ skills, 770+ MCP servers, and 2,500+ marketplaces by April 2026, with 110,000+ monthly visitors to the official plugin directory.^[58] The official anthropics/skills repo reached 125,000 stars with 14,600 forks — the most-starred AI tooling repository in the Claude ecosystem.^[22]

Official Anthropic Plugin Directory

GitHub: anthropics/claude-plugins-official. 18,000 stars, 2,200 forks, 310 commits.^[7] Languages: Python (31.6%), TypeScript (28.9%), HTML (19.5%), Shell (13%), JS (7%).^[7]

Plugin hierarchy: Skill (one SKILL.md file) → Plugin (plugin.json + skills/ + commands/ + agents/ + .mcp.json) → Marketplace (GitHub repo as plugin registry).^[44]

New plugin capabilities in 2026:^[44]

• LSP server support (.lsp.json): Real-time code intelligence — TypeScript, Python, Rust official plugins available
• Background monitors (monitors/monitors.json): Watches logs/files, delivers stdout as Claude notifications
• Default settings (settings.json in plugin root): Can set default agent (e.g., "agent": "security-reviewer")
• Plugin bin/ on PATH (v2.1.91): Executables in bin/ added to Bash tool's PATH while plugin active

Verification model: External plugins submit via form. "Anthropic Verified" badge for additional review. Security checks: credential exfiltration, destructive commands, pipe-to-shell patterns. Anthropic does not warrant third-party plugin behavior.^[7][43]

Key finding: The frontend-design skill reached 277,000+ installs by March 2026 — the most-installed single skill — demonstrating that design system enforcement via SKILL.md is the highest-demand use case in the ecosystem.^[22]

Official Anthropic Skills Repository

GitHub: anthropics/skills. 125,000 stars, 14,600 forks. Actively maintained.^[22]

"Skills 2.0" evolution: Skills now bundle instructions + scripts + templates + reference materials — a significant jump from text-only SKILL.md files. Community term: "Skills 2.0."^[22]

Notable skill categories: Document skills (docx, pdf, pptx, xlsx — production-grade, ships in Claude.ai); Development (testing web apps, MCP server generation); Creative (art, music, design); Enterprise (communications, branding).^[22]

Desktop Extensions (.mcpb) Format

Published June 26, 2025 (initial .dxt format). September 11, 2025: migrated to .mcpb (MCP Bundle) — legacy files still functional.^[17]

Format: ZIP archive with manifest.json + server/ + dependencies/ + icon.png. Server types: "node", "python", "binary".^[17] Claude Desktop includes built-in Node.js runtime. Python servers bundle own lib/ directory. OS keychain storage for API keys. Automatic updates via curated directory. Enterprise controls: Group Policy/MDM, extension blocklists, private directories, pre-installation approval.^[17]

MetaMCP — Unified MCP Aggregator

GitHub: metatool-ai/metamcp. 2,200 stars. Docker-based. MetaMCP is itself an MCP server — plugs into any MCP client.^[23]

Problem solved: 7 active MCP servers consume ~67,300 tokens at session start — over one-third of a 200K context budget. Lazy loading reduces context usage by up to 95%.^[34] MetaMCP provides the gateway to manage this via namespaces, tool picking, and multiple workspaces to prevent context pollution.^[23]

MCP Lazy Loading — Emerging Architecture

7 active MCP servers = ~67,300 tokens consumed at session start, exceeding one-third of a 200K context budget.^[34] Proposed lazy loading architecture from the community: lazy flag in settings.json per server; Skills/agents declare required MCPs in frontmatter (mcp: { required: [postgres], optional: [redis] }); MCPs load only when specific agents execute — not into main conversation thread. Proposed 95% context reduction.^[34]

Already shipping: Deferred tool schemas (ToolSearch capability) within sessions — schema-on-demand loading is functional. Full per-server lazy loading is an emerging pattern that may ship as official Anthropic feature.^[34]

Notable Community Plugins

Third-party marketplaces: SkillsMP.com (universal SKILL.md format, community ratings), SkillKit (400,000+ skills).^[48]

Section 7: Hook System Innovations (Q1–Q2 2026)

The hook system expanded from 12 documented events to 32 events in official docs by April 2026 — a 167% increase in the event surface area.^[8][53] Five handler types are now available, including three new types added in Q1 2026: HTTP, MCP Tool (v2.1.118), and Agent (experimental).^[8]

Complete Hook Event Reference (32 Events)

Source discrepancy note: Community guides (claudefa.st, Pixelmojo) list 12 events — these are older/partial references. Official docs (raw_8.md, raw_53.md) document 27–32 events as of April 2026. Always use official docs.^[8][45][52]

Five Hook Handler Types

Exit code semantics (IMPORTANT): Exit code 0 = success (allows action); Exit code 2 = blocking error (blocks action, stderr to Claude); Any other code = non-blocking (shows error, continues). Exit code 1 is NON-BLOCKING.^[8]

Key Hook Innovations Q1–Q2 2026

Key finding: The defer PreToolUse decision (v2.1.89) enables genuine human-in-the-loop patterns for CI/CD approval gates and multi-agent coordination — without polling. Combined with the MCP Tool handler type (v2.1.118), hooks can now orchestrate entire external service workflows from a single hook invocation.^[8][9]

MCP Elicitation (v2.1.76 — March 14, 2026)

MCP servers can pause execution and request structured input from users via elicitation/create. Two modes: Form (structured JSON schema validation) and URL (redirect to external URL for OAuth, payment flows).^[24] Hook integration: Elicitation and ElicitationResult events allow hooks to intercept and pre-answer elicitation requests programmatically.^[24] Security constraint: servers MUST NOT require PII, credentials, or sensitive data through elicitation — sensitive operations must use URL mode.^[24]

Hook Scope Levels

Community Hook Patterns (disler/claude-code-hooks-mastery)

3,600 stars, 602 forks. 13 demonstrated hook types.^[14] Key patterns:

• UserPromptSubmit as Firewall: Earliest intervention point — block dangerous prompts before Claude sees them, or inject project context^[14]
• Stop Hook Blocking: Stop hooks can FORCE Claude to continue — useful for task completion guarantees, but risks infinite loops^[14]
• UV Single-File Scripts Pattern: Hooks as standalone Python scripts with embedded dependencies via Astral UV — no shared venv management, fully portable^[14]
• PreCompact Backups: Fires before session compaction — community-validated transcript preservation pattern^[14]
• Team-Based Validation: Pairs builder (all tools) + validator (READ-ONLY) agents for async quality verification^[14]
• Meta-Agent Innovation: Agent that builds agents — specialized sub-agent that builds new agents from descriptions^[14]

Section 8: Claude Code Platform Releases Q1–Q2 2026

Claude Code shipped 5 major capability releases across Weeks 13–17 (March 23 – April 24, 2026): /ultraplan, Monitor Tool, /ultrareview, Routines, and Agent Teams, alongside a quality regression postmortem that revealed prompt changes have measurable intelligence impact.^[20][9]

Major Feature Release Timeline

/ultraplan — Cloud Planning System

30-minute remote planning sessions on Anthropic infrastructure running Opus 4.6. Local terminal polls every 3 seconds for updates. Browser-based review with inline comments + emoji reactions before execution. Plans can be "teleported" back to local working directory for execution. v2.1.101: First run auto-creates default cloud environment (no web setup needed).^[15][20]

/ultrareview — Cloud Multi-Agent Code Review

Public research preview since v2.1.86 (April 22, 2026).^[12]

Monitor Tool (v2.1.98)

Background watcher streaming events into conversation as transcript messages Claude reacts to immediately. Eliminates Bash sleep polling loops. Monitors: log files, CI/CD pipeline results, dev server crashes. Integration with /loop: Claude now self-paces — schedules next tick based on task, or uses Monitor to skip polling entirely.^[20]

Routines — Scheduled Cloud Agents

Saved Claude Code configurations running automatically on Anthropic-managed cloud infrastructure. Three trigger types (combinable): Schedule (hourly, daily, weekdays, weekly, or one-off timestamp), API (HTTP POST to per-routine endpoint with bearer token), GitHub (PR events with filter conditions: author, title, body, base branch, head branch, labels, is_draft, is_merged).^[21]

Usage limits: Pro = 5/day, Max = 25/day scheduled runs; GitHub webhook and one-off runs do NOT count toward limit.^[21] Repositories cloned fresh per run. Creates claude/-prefixed branches by default.^[21]

Agent Teams (Experimental)

Enable: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Requires v2.1.32+.^[63] Architecture: Team lead + Teammates (separate Claude Code instances, own context windows). Shared task list with dependency tracking + file-locking. Mailbox for direct agent-to-agent messages.^[63]

vs Subagents: Subagents report to main agent only; teammates communicate directly with each other.^[63]

Quality gate hook integration: TeammateIdle (exit 2 = send feedback, keep working), TaskCreated (exit 2 = prevent task creation), TaskCompleted (exit 2 = prevent task completion).^[63]

Limitations: No session resumption with in-process teammates, task status can lag, no nested teams, no per-teammate permission modes at spawn.^[63]

Quality Regression Postmortem (April 20, 2026)

Three simultaneous quality issues discovered and fixed in April 2026:^[9]

1. Default reasoning effort reduced from high → medium (March 4); reverted April 7
2. Caching bug (March 26) repeatedly cleared thinking history; fixed April 10
3. Verbosity prompt change (April 16) reduced coding quality; reverted April 20

Key finding: "Length limits: keep text between tool calls to ≤25 words" caused measurable intelligence drops. Lesson: prompt changes have detectable, measurable quality impact.^[9]

Key finding: The Monitor Tool and Routines together shift Claude Code from interactive-only to autonomous — agents can now self-schedule, react to external events (GitHub PRs, Sentry alerts via API trigger), and run nightly tasks without any human trigger.^[20][21]

Other Notable Q1–Q2 Releases

• Remote Control: Connect to active sessions from outside originating terminal — headless agent management (single controller currently)^[18]
• Dispatch: Trigger Claude Code programmatically as background worker jobs (webhook-triggered reviews, CI/CD pipeline steps)^[18]
• Channels: Real-time messaging/observability layer — streams tool calls, file changes, completion signals^[18]
• Opus 4.7 Vision: 2576px max on long edge, 98.5% visual acuity, high-res support for computer use/screenshots^[18]
• Computer Use in CLI (macOS): Open native apps, click through UI, test own changes from terminal. Requirements: Pro/Max plan, v2.1.85+, macOS only. Rolling visual context window cuts average action latency ~50%^[62]
• Rate Limits API (Apr 24): Programmatic query of org/workspace rate limits^[9]
• Claude Haiku 3 retired (Apr 20): Upgrade path to Claude Haiku 4.5^[9]

Section 9: Anthropic Platform Architecture & Roadmap

Anthropic launched Managed Agents public beta (April 8, 2026) with built-in filesystem memory (April 23, 2026), decoupled the agent Brain from Hands via durable server-side session events, and leaked plans for KAIROS persistent memory (May 2026), Coordinator Mode orchestration, and Daemon Mode persistent sessions.^{[13][49][61][15]}

Managed Agents — Core Architecture

Public beta launched April 8, 2026. Pricing: standard token rates + $0.08 per session-hour for active runtime.^[49]

Decoupled architecture (raw_13.md): Separates Brain (Claude + harness), Hands (sandboxes), and Session (event log stored OUTSIDE both). Events can be transformed in harness before passing to Claude (context organization + prompt caching optimization).^[13]

Performance from decoupling: P50 TTFT ~60% reduction; P95 TTFT >90% reduction.^[13]

Credential isolation: Two mechanisms — resource-bundled auth (embedded in remote URLs) + external vault + proxy pattern — prevent prompt injection from compromising credentials.^[13]

Built-in tools: Bash, file operations, web search/fetch, MCP servers, sub-agent spawning (research preview). Built-in prompt caching and context compaction (automatic).^[49]

Managed Agent Memory Beta (April 23, 2026)

Filesystem-based persistent memory — workspace-scoped Markdown files agents read/write/update across sessions.^[61]

• Files capped at 100KB per file (~25K tokens)^[61]
• Access scopes: org-wide read-only + per-user read-write + concurrent agent access without overwriting^[61]
• Audit logs: which agent, which session generated each memory entry^[61]
• Version control: rollback + content redaction^[61]
• Claude Opus 4.7 is recommended model: "most discerning about what to retain; produces most actionable entries at max effort"^[61]

Real-world results: Rakuten: 97% fewer first-pass errors, 27% lower cost, 34% lower latency. Wisedocs: 30% faster document verification.^[61]

Key finding: Anthropic's managed agent memory achieves Rakuten's 97% reduction in first-pass errors via filesystem-scoped Markdown that survives session resets — not a vector database, not a proprietary API: plain files with audit trails.^[61]

KAIROS — Persistent Cross-Session Memory (Leaked, May 2026 Launch)

Architecture from March 31, 2026 source map leak: daily append-only markdown logs at ~/.claude/.../logs/YYYY/MM/DD.md.^[15]

Auto-Dream Consolidation Cycle (4 phases):^[15]

1. Orient — scans recent session logs
2. Gather — extracts patterns and decisions
3. Consolidate — compresses to durable memory (<25KB output)
4. Prune — removes noise

Triggers after 24 hours AND minimum 5 sessions. 15-second blocking budget per phase before auto-backgrounding. Exclusive KAIROS tools: SendUserFile, PushNotification, SubscribePR, SleepTool. Teaser April 1–7, 2026. Full launch planned May 2026.^[15]

Coordinator Mode — Multi-Agent Orchestration (Partially Built)

One Claude as coordinator, spawning isolated worker instances with dedicated scratch directories. XML-based <task-notification> protocol: status, summaries, token usage, duration metrics. 4-phase workflow: research → specification → implementation → verification. Activation: CLAUDE_CODE_COORDINATOR_MODE=1. Status: requires additional development before release.^[15]

Other Leaked Features (March 31, 2026 — Unverified)

Controversial leaked features (treat as unverified): "Undercover Mode" (CLAUDE_CODE_UNDERCOVER) strips AI evidence from git commits; "Anti-Distillation" injects fake tools to poison competitor training pipelines; Frustration Detection monitors negative language and adapts tone.^[15]

MCP Ecosystem Foundation

MCP donated to the Linux Foundation's Agentic AI Foundation in December 2025. OpenAI and Google DeepMind adopted MCP in early 2025. Registry: 400+ community-built servers, top 50 averaging 12,000+ monthly installs each.^[50]

Claude Design (April 17, 2026): New Anthropic Labs product for collaborative visual creation of designs, prototypes, slides, and one-pagers. Separate from Claude Code but integrated into the Figma MCP loop.^[9]

Section 10: MCP Ecosystem Scale & Directory

The MCP ecosystem reached 70+ production-ready servers across 12 categories in 2026, with 10,000+ total MCP servers listed across directories (most are weekend projects).^[16] Remote/hosted MCP deployment emerged as the dominant Q1 2026 trend — Azure 2.0, VictoriaMetrics, Supabase, Sentry, and Jira all moved to remote OAuth, eliminating local installation.^[16]

Q1 2026 Major New Official MCP Releases

Trending Tools by Stars

Section 11: What Not to Reinvent

Based on the Q1–Q2 2026 ecosystem state, the following are definitively shipping — do not build custom alternatives:

Key finding: The 2026 Q1–Q2 release cadence compressed 2+ years of expected tooling gaps into a single quarter. Teams building custom observability wrappers, approval-gate systems, scheduled agent runners, or cloud review pipelines are duplicating infrastructure that Anthropic ships at a fraction of the cost and with direct platform support.^{[20][12][21][29]}

Section 12: Gaps & Missing Coverage

• Agentic testing MCPs beyond Storybook: No coverage of dedicated agentic test-generation MCPs (Testim, Mabl, Applitools) in the corpus. Storybook is the only confirmed agentic testing MCP.
• Pa11y MCP gap confirmed and unfilled: No Pa11y MCP exists as of April 2026. Community has not built one despite Pa11y's CI ecosystem maturity.
• Storybook MCP non-React support timeline: Vue, Angular, Web Components, Svelte support is planned but no dates published.^[27]
• KAIROS availability on Windows/MSYS: Daily append-only logs and UDS Inbox patterns assume Unix paths; Windows compatibility of KAIROS is unconfirmed in the corpus.^[15]
• Agent Teams stability timeline: No confirmed GA date for Agent Teams — experimental flag only, no roadmap signals in corpus.^[63]
• Routines rate limit increase: Pro=5/Max=25 scheduled runs/day is a sharp constraint for multi-project workflows — no roadmap signal for higher tiers.^[21]
• Coordinator Mode release date: Leaked but described as requiring additional development — no ETA in corpus.^[15]
• LSP server plugin pricing/performance: TypeScript, Python, Rust official LSP plugins exist but no latency or resource cost data in corpus.^[44]

Home