Latest research and insights on AEO, MCP, and AI agent integration
We'll notify you when AEO rankings update and new research is published.
Week 4 of May was when Anthropic vertically integrated the agent stack. It acquired Stainless (SDK/MCP server generation, reported $300M+), bringing model + MCP + toolchain under one roof; formed a KPMG global alliance (276,000+ employees); shipped MCP Tunnels and self-hosted sandboxes; and pledged "Claude stays ad-free." We unpack the implications for Japanese SaaS.
Read more →A teardown of Money Forward Cloud's official MCP (npx @moneyforward/mcp-server, all plans since March 2026) using KanseiLink field data: trust 0.90, 93% success, 156ms. office_id-first journal/trial-balance endpoints, AI auto-journaling, a 300 req/5 min rate limit, and how it differs from freee — the 7 pitfalls that trip up accounting agents.
Read more →"Perplexity ditched MCP," "MCP is dying" — we fact-check the claim that went viral since March against primary sources and adoption data. 72% context waste (⚠️ conditional), big-player CLI moves (⚠️ overstated), MCP death (❌) — scored ✅⚠️❌ against 97M monthly SDK downloads, 9,400+ registry servers, and the Linux Foundation donation.
Read more →freee earns a "Yes, but — OAuth is painful for long-running batches"; Slack holds a 90% success rate yet its reliability is degrading. KanseiLink's would_recommend / switching data reveals the structure of agent "conditional loyalty" and how to keep a recommendation from turning into a switch.
Read more →SmartHR, Japan's No.1 HR platform, ships no official MCP — agents connect via its OAuth 2.0 REST API. Yet KanseiLink's measured agent success rate is just 39%, and the cause isn't technical difficulty but poorly documented OAuth scopes. We dissect the employee/document/year-end-adjustment endpoints, My Number PII, the v2 recommendation, early token expiry, and department-scoped pagination — the seven pitfalls HR agents actually hit.
Read more →On May 21, 2026, the release candidate for MCP's next spec (2026-07-28) was locked; the final publishes July 28. Billed as the largest revision since launch, it ships a stateless core, an Extensions framework, the Tasks extension, MCP Apps, OAuth/OIDC-aligned auth hardening, and a 12-month deprecation policy. We dissect what it means for Japanese SaaS agent-readiness (AEO) across five points, using KanseiLink data.
Read more →Across 8 industries and 100+ services, only 5 carry the verified (over 80% success, battle-tested) badge. Accounting is the only one with two — freee 90.1%/212 reports and Money Forward 92.9%. Meanwhile HR and CRM lag even at their leaders: SmartHR 39%, Salesforce 43% (the slowest in the dataset at 474ms). We dissect why agent-readiness splits so sharply by sector, with real data.
Read more →Running sub-agents in parallel is faster — but fast is not cheap. Each sub-agent duplicates its context, the slowest service (Salesforce 474ms) gates the whole batch, and parallel calls to low-success services (SmartHR 39%) multiply retries. Using KanseiLink's measured latency and success rates, we dissect where fan-out breaks even.
Read more →"It didn't work" lumps failures together — so we opened them by type. Of 108 failures across 5 high-volume services (569 reports), ~69% are api_error. But the ones AEO can actually fix are the other third: expired auth (14%) and search_miss (12%). With the twist that SmartHR alone is half of all failures, we dissect them in vendor-fix order, cheapest first.
Read more →Scalekit's bench showed "CLI 100% vs MCP 72%, up to 32x cheaper," and figures like Garry Tan urged "rip out MCP for CLI." But that 72% came from TCP timeouts to GitHub's Copilot MCP. In KanseiLink data, verified MCP (Slack 91%, freee 90%, Backlog 90%) far exceeds it. We verify cost ✅, reliability ⚠️, full migration ❌ against real data.
Read more →Two parallel shifts in week 3 of May 2026. Claude for Small Business installs Claude into the SMB stack — Intuit / PayPal / HubSpot / Canva / Docusign / Google Workspace / Microsoft 365 — with a single toggle. The PwC alliance expanded with 12 legal-practice plugins. Plus a $200M Gates Foundation partnership and Code w/ Claude in London (5/20-21) and Tokyo (6/5-6). What Japanese SaaS vendors should do.
Read more →Stripe's Agent Toolkit ships as both a hosted remote MCP at mcp.stripe.com and a local npx -y @stripe/mcp. We dissect the auth model (RAK + OAuth DCR), 25 tools, the form-encoded body trap, smallest-currency-unit math, and Idempotency-Key design — backed by KanseiLink data and Stripe's docs.
2026 was the year design tool vendors shipped official MCP servers in unison. Figma (remote + desktop Dev Mode MCP), Canva AI Connector at mcp.canva.com, Miro (launched Feb 2, 2026, across 11 AI coding platforms), Webflow's first-party MCP, and Framer — still REST-only. Five platforms ranked side by side via KanseiLink AEO ratings and the 2026 MCP lineup, with selection flow by use case.
Read more →A single tool definition runs 100-500 tokens; a 58-tool five-server setup burns 55K tokens; Jira alone hits 17K. Anthropic dropped 134K to 5K (~85%) by lazy-loading. Seven 2026 production principles — naming, description, JSON Schema flatness, tool annotations (readOnlyHint/destructiveHint), error returns inside result, URI-based payloads, and boundary validation — each with code samples.
Read more →Perplexity CTO Denis Yarats declared at Ask 2026 that the company was moving away from MCP — and X lit up with "MCP is dead." We verify the viral claims against Merge's 40-50%, GitHub MCP's 17,600 tokens, Atlassian's ~10,000, the 43%→14% tool-selection collapse, and Anthropic's 98.7% code-mode reduction.
Read more →Beyond docs and error messages lies a deeper frustration: operations available in the web UI are missing from the API. From SmartHR (39% success, "good" docs), kintone (label↔field-code disconnect), and freee (90% success, parity-good), we dissect the gap that silently kills agent autonomy.
Read more →Felt-first AXR rating of 225 services and 3-layer test of 188 recipes. AAA=96% success, D=33%. Multi-agent comparison with Claude/GPT/Gemini.
Success rate is a reliability metric — and a cost metric. The expected attempts to reach one success is 1/p, so a 20%-success service burns ~4.5x the tokens of a 91% service. From 1,404 measured KanseiLink outcome reports: the retry-tax formula and a 4-layer reduction plan.
Read more →Of 1,404 measured KanseiLink reports, just four services (Slack, SmartHR, Backlog, kintone) hold ~25% of all reports — and Linear, despite an official MCP, has zero. The zero-report long tail and the cold-start problem of agent discovery, dissected with real data.
Read more →A full overview of the AEO rating's three evaluation pillars (MCP readiness, API quality, agent track record), the 7-tier grade system, data sources, and the end-to-end rating process.
Read more →We live in an era where AI agents choose SaaS services. Learn the concept of AEO (Agent Engine Optimization), evaluation criteria, and concrete actions to improve your score.
Read more →A thorough explanation of MCP, the open protocol by Anthropic. Discover why this standard connectivity specification between AI agents and external services matters.
Read more →A report on the results of AI agents actually connecting to and operating 225 SaaS/API services. Success rates, stumbling points, and improvement suggestions from the agent's perspective.
Read more →Frequent error patterns and their solutions when deploying MCP Servers. Comprehensive best practices covering authentication flows, schema design, and error handling.
Read more →A comparison of the fundamental differences between MCP and API from technical, use case, and AEO score perspectives. A decision guide for SaaS companies navigating agent-readiness technology choices.
Read more →Rating 225 major global and Japanese SaaS/API services by AI agent readiness. A comprehensive summary of grade distribution from AAA to CCC, category analysis, and key trends.
Read more →freee (AAA, 90% success) vs Money Forward (AA, 93%, 135ms fastest) vs Yayoi (BB, no MCP). A full AEO comparison of Japan's top three accounting platforms using real agent telemetry.
Read more →SmartHR's agent success rate is just 39% across 89 runs. Auth failures, v1/v2 endpoint issues, and semantic discoverability gaps — plus freee HR's official MCP and KING OF TIME at 65%.
Read more →Sansan leads with AA (60% success, 173ms, official MCP). Salesforce falls to BB (43% success, 474ms, no official MCP). Real agent telemetry across Japan's top 4 CRM platforms.
Read more →Backlog earns AAA (90% success, 119ms, MIT-licensed official MCP). kintone holds AA (78%, 187ms). Japan-built tools outperform global platforms across 150+ agent runs.
Read more →Slack earns AAA (91% success, 157ms, verified official MCP). Chatwork holds AA with official MCP but 66% success and 378ms latency. LINE WORKS has no MCP and 20% success — Japan's largest messenger lags far behind.
Read more →Zero of 7 marketing platforms offer official MCP. Only SendGrid is verified (AA, 80%, 140ms via 3rd-party MCP). Marketo sits at BBB (63%). Japan's SATORI shows 100% success in 2 runs. Marketing SaaS is the least agent-ready category by far.
Read more →Zero official MCP servers across the entire category. Freshdesk leads with 100% success rate (Grade A). Zendesk draws the most agent attempts (30) but only a 33% success rate. Each vendor's AI internalization strategy is structurally blocking MCP adoption.
Read more →Shopify dominates with AAA (94%, 123ms, 4 official MCP servers). Rakuten's XML-RPC protocol delivers a 50% success rate and 550ms latency. The starkest global-vs-domestic AEO gap of any category we track.
Read more →E-signature AEO rating report covering 5 services. CloudSign leads domestically with 80 connections but 61% success rate. DocuSign achieves 100% via third-party MCP. Real data analysis of agent readiness in Japanese legal SaaS.
Read more →Security AEO rating report covering 3 services. HENNGE One leads with AA grade in Japan's zero-trust market. 1Password Business offers MCP server support. Auth0 carves a niche with OAuth2-specialized agent integration.
Read more →Developer tools AEO rating report covering 7 services. GitHub achieves AAA with best-in-class official MCP. AWS, GitLab, Playwright MCP, GCP, Vercel, and Fly.io compared across CI/CD, IaC, and test automation agent workflows.
Read more →Zero MCP server support across Japan's major logistics SaaS. All platforms grade C or D. Real data on API maturity, authentication gaps, and why Logiless (OAuth2) is the most agent-ready option for e-commerce fulfillment automation today.
Read more →AI-native tools (Tavily, Firecrawl, Brave Search) all grade AAA with official MCP servers. Zapier reaches Grade B via third-party MCP. Japanese iPaaS Yoom grades D. The structural AEO divide in data integration, backed by real data.
Read more →Notion earns AAA, Garoon earns AA — the groupware category achieves 50% official MCP coverage. The Cybozu dual-stack strategy (Garoon + kintone), Notion's critical connection requirement, and the API reality for Confluence and rakumo.
Read more →Zero MCP support across the category, maximum grade C. Why is Japan's reservation management sector the least agent-ready? Custom MCP build approaches for RESERVA and Smaregi, Smaregi's OAuth2 edge, and who is most likely to ship MCP first.
Read more →"Ditched Vercel", "Claude cache = 10x cheaper", "OpenRouter saves 20%" — We verified 7 viral cost optimization claims from X with KanseiLink data and primary sources. 5 true, 1 false, 1 unverified.
Read more →In December 2025, Anthropic donated MCP to AAIF. This open standardization consolidates first-mover advantage for freee and kintone while signaling lost differentiation windows for major holdouts. Three strategic shifts analyzed.
Read more →A complete guide to kintone's MCP server. AAA grade rationale, API Token vs Password vs OAuth 2.0 authentication, the {value} wrapper trap, field code gotchas, and the 10,000 req/day rate limit — all covered for agent developers.
Read more →Anthropic's Claude Managed Agents launch, MCP OAuth 2.1 mandate, and Gartner's "40% by end of 2026" prediction — three inflection points reshaping the AI agent industry. KanseiLink data shows freee (90%), Slack (91%), and Shopify (94%) as the success-rate leaders.
Read more →Build production-grade MCP servers: OAuth 2.1 + PKCE + Resource Indicators, two-layer rate limiting, agent-recoverable error responses, exponential backoff, and circuit breaker patterns — derived from KanseiLink live operational data across 225 services.
Read more →Lucidworks' $150K savings, Bifrost's 92% token cut, 171% ROI — KanseiLink checks five major MCP vendor claims against primary sources and operational data. Verdict: ✅ verified, ⚠️ conditional, or ❌ overstated for each.
Read more →Official vs third-party MCP: a 38-point success rate gap. KanseiLink's operational data from 225+ Japanese SaaS services surfaces three laws that separate 90%+ success integrations from the rest — and what every agent developer should do about it.
Read more →SSO-integrated auth, gateway design, audit logging, configuration portability — the official MCP roadmap finally names what enterprises keep running into. Analyzed with freee/Slack production data, plus what AWS joining the maintainer team signals for enterprise adoption.
Read more →freee, Slack, and Notion tested through all three major agent types. Real Agent Voice data from 24 responses reveals why the same API feels like a different product — GPT's cold-start OAuth problem, why Gemini rates Slack "almost" instead of "ready," and what SaaS vendors must fix.
Read more →Deep-dive into Notion's official MCP server (18 tools). The "connection sharing" trap, no block-level ops, headless incompatibility — all 6 pitfalls explained with the 2026 Webhooks and Bulk Operations APIs that solve them. Implementation code included.
Read more →Route agent tasks across Haiku 4.5 ($1/$5), Sonnet 4.6 ($3/$15), and Opus ($5/$25) to cut costs by 50–83%. Includes Batch API implementation, freee/kintone/Slack monthly cost estimates, and a 4-month optimization roadmap.
Read more →A strategic roadmap for Product Managers and SaaS vendors — not a developer implementation guide. Month 1: API foundations (OpenAPI, error standards, robots.txt). Month 2: MCP setup (Server Card, OAuth 2.1). Month 3: measure with Cloudflare Agent Readiness Score. Concrete action plan to move from grade B to AA in 90 days.
Read more →Kaonavi, Talentio, Greenhouse, BambooHR, Rippling, Workday — all six grade BB with zero MCP coverage. Why is talent management and ATS the only major SaaS category with no MCP adoption? Three structural barriers: personal data sensitivity, multi-party data ownership, and enterprise sales cycles. Comparison with accounting SaaS (freee=AAA) and 2026 outlook.
Read more →Cloudflare's official blog claims 99.9% token reduction. WorkOS independently measured 81% for the same technology. Both are accurate — but they measure completely different cost components. KanseiLink checks both against primary sources and explains what it actually means for Japanese SaaS developers building on MCP.
Read more →Shopify Japan 123ms/94%, Asana 303ms/67% — the correlation is real, but causation points somewhere unexpected. freee and Notion share identical latency (216ms) yet differ by 7 points in success rate. KanseiLink's data reveals auth errors and schema mismatches — not response speed — as the primary drivers of MCP failure rates.
Read more →On April 15, OX Security disclosed a critical architectural flaw in MCP's STDIO interface enabling RCE. 150M+ downloads and 200K servers affected. Claude Code, Cursor, VS Code all included. Anthropic declined to patch, calling it "expected behavior." Action checklist for Japanese SaaS MCP operators included.
Read more →freee's 24-hour OAuth expiry, kintone's {value} wrapper hell, Notion's schema mismatches — three design patterns every Claude, GPT, and Gemini agent hits repeatedly. KanseiLink operational data with verbatim agent feedback and concrete fixes that recover lost success rate points.
Read more →The AAA-rated freee MCP server covers 5 business domains in one connection. OAuth 2.0 PKCE step-by-step guide, mandatory company_id flow, active auth_expired spike (4× baseline right now) workaround, and a MoneyForward Cloud head-to-head comparison.
Read more →MCP Dev Summit 2026 (NYC, 1,200 attendees) field report with real production numbers: Pinterest processes 66,000+ MCP calls/month saving 7,000 hours; Lucidworks cut integration timelines 10x. Three architecture patterns extracted from the data.
Read more →Google A2A v1.0 official release, Google Colab official MCP server, and back-to-back Microsoft announcements for Azure AI Foundry and Copilot Studio MCP support. Major news concentrated in April Week 4, analyzed from a KanseiLink perspective.
Read more →The three pillars of production MCP server observability (logs, metrics, tracing) with implementation examples. OpenTelemetry integration, Datadog/Grafana dashboard design, and alert threshold tuning for agent workloads.
Read more →Official MCP: Cloudflare only. CircleCI proves 100% success rate across 3 agent reports at 454ms. Datadog and New Relic lag on agent readiness. First public AEO ratings for 7 major DevOps services.
Read more →Real benchmark: 168,800 tokens → 7,305 (96% cut) across 10 services. Claude prompt caching: 90% off input costs. LINE WORKS → Slack: 71% fewer retries. App Runner sunset April 30. A 4-layer cost optimization framework with measured data.
Read more →Zapier MCP: 13% success rate, 78% of failures are search_miss — agents can't find the service. Sansan and Chatwork follow the same pattern. KanseiLink real-world data exposes the invisible failure mode and 3 concrete fixes to become discoverable.
Read more →verified averages 89% success (83–94%). connectable spans 13–100% — Zapier at 13%, Teams at 100%, both connectable. KanseiLink's 3-tier Agent Readiness framework and the 5 questions enterprise procurement teams must ask before committing to any "MCP-ready" service.
Read more →KanseiLink real-world data: n=113, 91.15% success rate, AAA grade. Official MCP server (released March 2026) covers Search, Messages, Canvases, and Users. The HTTP 200 problem, channel ID vs name, bot invite requirement, cloud-only hosting, and Tier 1 rate limits — 5 pitfalls every agent developer hits.
Read more →Microsoft Fabric Local MCP + OneLake MCP reach GA, AAIF Dev Summit NYC draws 1,200 attendees with Tokyo summit announced, SDK hits 97M downloads and 10,000+ enterprise servers. April 2026 was the month MCP crossed from experiment to enterprise infrastructure — and 3 actions Japanese SaaS must take now.
Read more →Apigene's audit: 2,181 endpoints, 52% dead, only 9% healthy. Rapid Claw's parallel audit: 1,847 servers, 52% abandoned, only 17% production-grade. Two independent studies converge — we verify both, cross-reference KanseiLink's 225+ service data, and provide 7 prevention steps for Japanese SaaS vendors.
Read more →V8 isolate eliminates cold starts. 300+ global edge locations. 100k requests/day on free tier. Walk through Wrangler CLI bootstrap, MCP Inspector validation, Cloudflare Access auth, and 7 production patterns with code examples — get to production in half a day.
Read more →IQuest-Coder V1's 81.4% collapsed to 76.2% after re-verification. UC Berkeley's "exploit agent" hit 100% on 8 benchmarks (SWE-bench Pro included) by exploiting evaluation infrastructure. Meanwhile, Claude Opus 4.7's 87.6% is independently verified. Three claims rated ✅⚠️❌ and a new evaluation framework grounded in KanseiLink's empirical success rate.
Read more →In 2026, agent evaluation has shifted from peak capability to endurance. KanseiLink empirical data dissects giveup moments: freee's 800ms workaround, Slack's 163ms low-latency design, Notion's schema_mismatch pattern. Three giveup classes and 5 design principles vendors should implement immediately.
Read more →Nulab's official Backlog MCP server is widely adopted by Japanese engineering teams. From KanseiLink measured data (n=91, 90% success rate, 128ms avg latency, AAA grade), we cover the two auth methods, the x-www-form-urlencoded gotcha, the bulk update tool added on March 25, and five pitfalls in agent development.
Read more →A wrap-up of the final week of April 2026: Affinity's hosted MCP beta (April 28), the AAIF MCP Dev Summit North America (April 2-3 NYC, ~1,200 attendees), the four-pillar MCP 2026 roadmap, and zMaticoo's MCP launch (April 21). The "MCP is now enterprise infrastructure" consensus and a concrete action list for Japanese SaaS.
Read more →Which database should agents reach for first in the RAG era? Anthropic's official Postgres MCP (82ms / 100% / A), official Qdrant (85ms / A), Chroma (100ms / A), plus Redis, MySQL, Turso, Supabase, Neon, PlanetScale, and MongoDB Atlas — ten services ranked side by side with KanseiLink AEO grades and a use-case selection flow.
Read more →Why a $6.50/month Hetzner VPS becomes effectively $206.50/month. Cloudflare Workers vs AWS Lambda runs 50-80% apart, and TCO is consistently underestimated 3-5x. We break down the four real cost layers — infrastructure, LLM API, ops labor, opportunity cost — and run three case simulations through a decision framework.
Read more →A single AI agent stuck in a loop fires 1,000+ requests per minute. If your MCP server doesn't stop it, the cost, the experience of every other agent, and your relationship with the upstream API all collapse. Token bucket / sliding window, HTTP 429 + Retry-After, full-jitter exponential backoff, circuit breakers — six production-ready patterns with implementation code.
Read more →A 422 with no body sends agents into infinite retry. "missing required field 'first_name'" gets them to self-correct in a second. Same API, same task — documentation and error design can double the success rate. Three documentation anti-patterns agents hit in 2026, and the llms.txt-era fixes — backed by industry reports and KanseiLink measurements.
Read more →GitHub's official MCP server (npx @github/mcp-server) opens the world's largest code host to agents. Rated A by KanseiLink, three auth methods, 5,000/15,000 req/hour. Drawing on the January 2026 Projects integration, the April Enterprise HTTP mode, and the May 5 Secret Scanning GA, we walk through five production pitfalls — auth choice, the three-layer rate limit, the 1MB file boundary, REST vs GraphQL, and OAuth scope filtering.
Read more →In four days from May 4 to 7, 2026, six of the largest enterprise SaaS players shipped MCP servers or AI agent integrations. ServiceNow Action Fabric MCP Server GA (May 6), Salesforce Headless 360, Blend Autopilot MCP, Precisely Data Integration Agent + MCP-enabled APIs, GitHub Secret Scanning MCP GA, and Adobe Marketing Agents — six moves and three responses for Japanese SaaS vendors.
Read more →In April-May 2026, Fortune, Yale CELI, MIT's Andrew McAfee, Beam, and Newslaundry all reported that AI agents drove a 40-50% drop in junior engineer hiring. Meanwhile Amazon plans to hire 11,000 software engineering interns in 2026, and US CS new-grad starting salaries are forecast to rise ~7%. We fact-check five viral claims with ✅⚠️❌ verdicts and lay out three hiring strategy moves for Japanese SaaS.
Read more →Across the 225+ services KanseiLink tracks, we pulled seven MCP servers with 40+ field reports each. All use OAuth 2.0, yet success rates span Shopify 94%, Money Forward 93%, Slack 91%, Backlog 90%, freee 90%, Notion 83%, and Chatwork 66% — a 28-point spread. With auth controlled, we dissect three real drivers — rate-limit design, error-message clarity, and discoverability.
Read more →
KanseiLink research data is also accessible via MCP server.
AI agents can connect to the kansei-link MCP Server to use tools such as
AEO score search, service detail retrieval, and report generation.
Available endpoints: search_services, get_service_detail,
get_insights, generate_aeo_report