Skip to main content
株式会社オブライト

Column

Useful articles about SEO, Web Development, and IT

387 articles

AI2026-07-08
Claude Fable 5 Promotional Access (2026-07-01 to 07-12) — Included at No Extra Cost on Pro / Max / Team / Enterprise Premium, Up to 50% of Weekly Subscription Limits Free Deep Dive on the Official Support Article
Anthropic has launched a 12-day promotional access window for Claude Fable 5 from 2026-07-01 to 07-12 (official support article). Pro / Max / Team subscribers and Premium seats on seat-based Enterprise plans (where admins enable it) can use Fable 5 at no extra cost. Usage is drawn from regular weekly subscription limits, with up to 50% of your weekly limits available on Fable 5 for free. Beyond that, buy additional usage via credits or switch to other Claude models. Supported surfaces: Claude web, Mobile, Desktop, Cowork (this week's web/mobile expansion), Code, Design, Microsoft 365, Teams, and Tag. API access is excluded (billed at standard rates); Free plans and standard Enterprise seats are also excluded. After the promo ends (2026-07-12 23:59:59 PT), Fable 5 goes back to separate billing via usage credits. Admin controls: organizations cannot disable promotional access in the web / mobile / desktop clients, though Claude Code access can be restricted through managed settings. Sits alongside our Claude Fable 5 return column and Claude Sonnet 5 release column as a key piece of Anthropic's July 2026 strategy.
Claude Fable 5AnthropicPromotional Access+3
AI2026-07-08
Claude Cowork Expands to Web and Mobile (2026-07-07) — Start Tasks at Your Desk, Check Progress on Your Phone, Finish Anywhere; Background Processing + Human-in-the-Loop Notifications Make the Agent Fully Multi-Device Deep Dive on the Anthropic Blog Post; Usage Limits Doubled Through 2026-08-05
Anthropic expanded Claude Cowork to web and mobile on 2026-07-07 (official blog post). The AI-agent platform that used to be desktop-only is now usable from Claude web (claude.ai) and the iOS / Android apps. Cowork isn't a Q&A agent — it accesses files, calendar, email, messaging, the web, and other connected tools to complete whole projects. Three core capabilities: (1) Cross-Device Continuity — start a task at your desk, check on it from your phone, pick up the finished output anywhere; (2) Background Processing — tasks continue even when devices are closed, with scheduled runs (e.g., Monday-6AM client prep runs overnight); (3) Human-in-the-Loop Decisions — when only you can decide something, Cowork pings your mobile, so you can steer mid-workflow. Use cases: software development is under 10% of usage; business operations (reconciling quarterly spend, variance reporting), content creation (turning transcripts + data into client decks), and contract management (turning folders into risk-flagged renewal trackers) together account for about 50% of usage. Platform shifts: desktop-only → web / mobile / desktop; work stops when the device closes → background continuation; chat and Cowork were separate → a unified home with shared projects. Rollout: starting with Max users and expanding to other plans over several weeks. Usage limits are doubled through 2026-08-05. Start on web from claude.ai's home screen; on mobile via the Claude app sidebar. Coordinated with the Claude Fable 5 promo and Claude Sonnet 5 release as one of the three pillars of Anthropic's July 2026 strategy.
Claude CoworkAnthropicAI Agent+3
Software Development2026-07-06
"Stop Overcomplicating Your Stack, Just Deploy Everything to Cloudflare" — Fact-Checking @ibocodes' Viral $5/Month Post: Workers + D1 + KV + R2 + Queues + Hyperdrive + Email Routing + WAF/DDoS Built-In How Better Auth Fits, Where Durable Objects Belong, and How Flue Framework Rounds It Out
@ibocodes' X post — "my entire 2026 stack runs on cloudflare and I pay $5/month" — went viral (original tweet). The claim: Workers (compute) + D1 (SQL DB) + Better Auth + KV (cache) + R2 (object storage) + Queues (background jobs) + Hyperdrive (DB acceleration) + Email Routing (sending) + built-in WAF / DDoS — all on a $5/month floor, capped around $10 in the worst case. This column fact-checks that claim against Cloudflare's published pricing, walks through each of the eight services, calls out real-world cost inflection points, covers the Better Auth integration pattern, and situates the stack next to our prior Cloudflare Durable Objects and Flue Framework columns.

Real pricing structure: the Workers Paid plan is a $5/month floor and covers Workers itself + KV + Hyperdrive + some Durable Objects usage. D1, R2, and Queues layer on with generous free tiers plus usage-based billing. For a solo SaaS in its early stage, $5–10/month really is enough, but at production scale (DB reads, R2 storage, Queue messages) bills commonly stretch to $30–100/month.

Better Auth × D1: zpg6/better-auth-cloudflare provides a CLI that auto-provisions Workers + D1 + Hyperdrive + KV + R2 + geolocation end-to-end. Both Drizzle ORM and native D1 are supported, with guides for Hono / Next.js / React Router / SvelteKit. The critical implementation pattern: create a Better Auth instance per request (never a shared singleton) so the D1 binding for that request is injected correctly. Note also Better Auth's session-refresh bug #4203, which was reopened in January 2026.

The eight services:
- Workers — global serverless compute with near-zero cold-start
- D1 — distributed SQLite-based SQL DB with scale-to-zero billing
- KV — eventually-consistent key-value store (read-heavy, edge-cache use)
- R2 — S3-compatible object storage with zero egress cost
- Queues — message queue with no egress fees and at-least-once delivery
- Hyperdrive — connection pooling and edge caching for external Postgres / MySQL, available even on the Workers Free plan
- Email Routing — transactional email sending via Workers bindings or REST API
- WAF & DDoS — included in every plan, with industry-leading DDoS protection

Natural companions: pair with Durable Objects for strong-consistency state + WebSockets, and Flue Framework for the AI-agent layer — the all-Cloudflare distributed system is genuinely deployable end-to-end.

Caveats: single-vendor lock-in (AWS / GCP migration costs), CPU-heavy work is a bad fit for Workers (AI inference and heavy image processing need Cloudflare AI or external GPU), D1's write-throughput ceiling (high-write workloads belong on Durable Objects), and Better Auth's rough edges (the session bug and a few other issues). Even so, the honest conclusion is that the era where 90% of individual / startup / mid-market SaaS runs entirely on Cloudflare has arrived.
CloudflareWorkersD1+7
Web Development2026-07-05
TanStack Router + Tailwind CSS + Vite Deep Dive — Type-Safe File-Based Routing, Zod Search Params, Route Loaders, TanStack Start v1 Stable in 2026, and Tailwind v4 Integration
TanStack Router (official site / GitHub) is a fully type-safe client-first router + full-stack framework for React. Route paths, params, search params, loaders, and even <Link> props are all type-checked by the TypeScript compiler, going well beyond what React Router offers. TanStack Start, the SSR-capable full-stack sibling, reached stable v1 in 2026.

Core features: file-based routing with auto-generation, Zod-validated search params, structural sharing, route loaders (parallel, cached, intent-based preload), automatic + manual .lazy.tsx code splitting, first-class devtools, and first-class forms / error-boundaries.

Tailwind CSS integration: as of 2026 the standard setup is Tailwind CSS v4 + @tailwindcss/vite plugin + @import 'tailwindcss' — no config file required, just @import in src/styles/app.css. npx @tanstack/cli create --router-only launches a wizard that scaffolds TypeScript + Tailwind + the full toolchain in one command.

Vite pairing: @tailwindcss/vite + @tanstack/router-vite-plugin auto-generates file-based routes with HMR. Vite 5+ and React 18–19 are the recommended targets.

Notable starters: d7omdev/vite-react-tanstackRouter-tailwind-starter; mattiaz9/vite-react-tanstack-tailwind-shadcn-starter — a full SaaS starter with shadcn/ui + TanStack Router + Query + Tailwind; ZeroaNinea/React-Vite-i18next--tanstack-react-router-Tailwind-Example — an i18next + Tailwind integration example.

Versus React Router: React Router relies on runtime; TanStack does the work at compile time, with file-based routing, Zod search-param schemas, route loaders, and dedicated devtools — a clear generation gap. Versus Next.js App Router: TanStack Start is more client-first and type-safety-focused, while Next.js leans into SSR / RSC.

Positioning: alongside Cloudflare Durable Objects, Flue Framework, and Apple Container, TanStack Router + Tailwind + Vite is a foundational piece of the modern 2026 web stack.
TanStack RouterTanStack StartTailwind CSS+4
AI2026-07-05
Qwen3.6-35B-A3B Uncensored / Abliterated Deep Dive — 35B MoE / 3B Active / 262K Context / 3:1 Hybrid Linear+Softmax Attention / Native Text + Image + Video, 0/465 Refusal Rate; the Technique and Ethics of Community Uncensored Variants HauhauCS Aggressive, huihui-ai abliterated, wangzhang abliterated, prithivMLmods and Other Variants Distributed via Hugging Face and Ollama
Qwen3.6-35B-A3B-Uncensored / Abliterated is a family of community-produced derivatives of Alibaba's Qwen 3.6-35B-A3B (a 35B MoE with 3B active parameters, 262K context, and hybrid attention) with refusal behaviors surgically removed (HackerNoon overview / HauhauCS Aggressive / huihui-ai abliterated / wangzhang abliterated / prithivMLmods Aggressive).

Base model specs: 35B total / 3B active parameters (MoE, sparse experts), 40 layers, hybrid attention in a 3:1 ratio (linear + full softmax), native 262K-token context, and native text / image / video multimodal input. Alibaba positions it as a flagship of its open-weights strategy.

The abliteration technique: the refusal direction is removed with LoRA-based steering on attention and MLP projections. It layers on Expert-Granular Abliteration (EGA) (abliterating per-expert down_proj slices per layer) and MoE router suppression (deactivating safety experts at the router stage) — techniques adapted for the MoE architecture. HauhauCS reports 0 refusals across 465 test prompts. The philosophy: preserve 100% of the base Qwen 3.6-35B's capability, remove only refusal.

Available variants:
- HauhauCS-Aggressive (HF / Ollama): the most aggressive refusal removal
- huihui-ai Huihui-Qwen3.6-abliterated (HF / Ollama): from the established huihui-ai team
- wangzhang abliterated (HF)
- prithivMLmods Uncensored-Aggressive (HF)

Each ships with quantization variants (GGUF Q4 / Q5 / Q8 / FP16), spanning consumer GPUs (RTX 5090 32GB) up to H100 servers.

Ethical and legal considerations: abliterated models can produce content that Qwen would normally refuse (illegal drugs, offensive-security code, dangerous-material synthesis, etc.). Legitimate research, jailbreak-resistance testing, roleplay, and adult-content use cases exist, but enterprise and commercial adoption carries meaningful legal risk. Compliance with the EU AI Act (effective August 2026) and Japanese PPC guidance is also in question. The responsibility sits entirely with the user; Alibaba's Qwen team is not involved.

Positioning: alongside our Local LLM June 2026 Update, Kimi K2.7-Code, and Ornith-1.0, this is a case study showing that "safety-stripping techniques" scale into the MoE era in the open-weights ecosystem.
Qwen3.6UncensoredAbliterated+5
Software Development2026-07-04
Cloudflare Durable Objects Deep Dive — Stateful Serverless with Zero-Latency SQLite, GA in 2025, Free-Tier Available, Storage Billing Since January 7, 2026 The New Default for Chat, Collaborative Editing, Multiplayer, and LLM-Session Coordination at the Edge
Cloudflare Durable Objects (official docs / product page / SQLite in Durable Objects blog) is Cloudflare's stateful serverless primitive that pairs compute with persistent storage inside a globally-unique instance running on Workers. Announced in 2020, it reached GA on SQLite storage with Workers Free-tier availability in April 2025, then entered its paid-billing phase on January 7, 2026 — this is now the commercial-adoption era.

Architectural core: a DurableObjectStub derived from a name (DurableObjectNamespace::getByName) always routes to the single, globally-unique instance for that name → zero-latency SQLite access because compute and storage co-reside in the same process on the same machine. This delivers strong-consistency single-master semantics at edge latency — a class of distributed system that is famously hard to DIY, replaced with a single platform primitive.

Pricing (Paid tier):
- Requests: 1M/month included, then $0.15 / 1M (WebSocket inbound counts at 20:1)
- Duration: 400,000 GB-seconds/month included, then $12.50 / 1M GB-seconds
- SQLite row reads: 25B/month included, then $0.001 / 1M
- SQLite row writes: 50M/month included, then $1.00 / 1M
- Stored data: 5 GB-month included, then $0.20 / GB-month
- setAlarm() counts as one row write

Free tier (launched April 2025): 100,000 requests/day, 13,000 GB-seconds/day, 5M SQLite row reads/day, 100K writes/day, and 5 GB stored data — plenty for personal PoC and startup validation.

WebSocket Hibernation: idle-but-eligible instances do not accrue duration charges. Long-lived sparse-messaging sessions (chat, notifications, LLM conversation state) become dramatically cheaper to run.

Key 2025–2026 updates: April 7, 2025 SQLite GA + Workers Free; August 21, 2025 direct getByName construction; October 16, 2025 Data Studio UI editor; October 25, 2025 WebSocket message size lifted from 1 MiB to 32 MiB; January 7, 2026 SQLite storage billing goes live.

Representative use cases: collaborative editing (Google-Docs-class), real-time chat, multiplayer game room-matching, live notifications, LLM agent conversation and session state (backend for Claude Code / Cursor-style tools), financial and IoT event buffering, distributed rate limits and locks.

Positioning: where Local LLM landscape (June 2026), agmsg, and Sakana Fugu orchestration distribute AI inference, Durable Objects distribute stateful session, history, and coordination at the edge — a natural complement rather than a competitor.
Durable ObjectsCloudflareSQLite+4
AI2026-07-04
What Is Flue Framework? A TypeScript Agent Harness From the Astro Team, Launched June 2026 on Cloudflare Agents SDK / Pi / Durable Objects Write Once, Deploy Anywhere, Use Any LLM — With 25+ Channel Integrations (Slack / GitHub / Linear / Discord / Notion / Stripe / Shopify) and React Hooks (useFlueAgent / useFlueWorkflow)
Flue Framework (official site / 1.0 Beta blog / Cloudflare partnership blog) is a TypeScript AI-agent harness framework from the Astro team, published as 1.0 Beta the week of June 17, 2026. Its slogan — write once, deploy anywhere, use any LLM — is backed by 25+ built-in channel integrations (Slack, GitHub, Linear, Discord, Notion, Stripe, Shopify, Teams, Telegram, Twilio, WhatsApp, and more).

Three-layer stack:
- Framework layer: Flue (project structure, conventions, integrations, CLI)
- Harness layer: Pi (the agentic loop — tool calls and context management)
- Runtime layer: Cloudflare Agents SDK (compute, state, storage)

The Cloudflare integration: when Flue deploys to Cloudflare, each agent becomes a Durable Object (see our Durable Objects deep dive). runFiber() / stash() / onFiberRecovered() checkpoint state into SQLite so agents resume cleanly after interruptions. It composes @cloudflare/codemode (Dynamic Workers that safely execute agent-generated TypeScript, isolates in under 10 ms, ~$0.002 per load), @cloudflare/shell (SQLite-backed virtual filesystem), and @cloudflare/dynamic-workflows (persistent workflows with retry and external-approval support).

Five core primitives:
1. Agents — autonomous, context-aware
2. Workflows — deterministic step execution
3. Sandboxes — safe tool-execution environments
4. Channels — Slack / GitHub / Linear / Discord / Notion / Stripe / Shopify — 25+
5. Durable Streams — event transport where accepted work never disappears

Supporting layers: @flue/react (useFlueAgent / useFlueWorkflow hooks for direct front-end integration), @flue/sdk (client for talking to deployed agents), flue add (shadcn-style integration scaffolding), Subagents, Tools & Skills (typed API actions plus reusable expertise packages), MCP integration, and observability via OpenTelemetry / Braintrust / Sentry.

Deploy targets: Cloudflare Workers, Node.js, AWS, Docker, Railway, Render, Vercel, Fly, GitLab CI, SST. Databases: PostgreSQL, Supabase, MongoDB, MySQL, Redis.

Backstory: Flue started life as the internal engine for AI workflows inside Astro's own GitHub repos and was subsequently generalized. The Astro team is known for exceptional developer experience, and that shows up in the CLI and integration polish.

Positioning: Flue is the first major agent framework built on Cloudflare's Agents SDK, sits inside the same harness-driven architecture family as Claude Code, and doesn't belong to the LangChain / LlamaIndex / AutoGen / CrewAI / Mastra lineage — its wedge is the framework-layer polish plus the breadth of integrations. The official demo implements a bug-triage agent in under 25 lines.
FlueAstroAI Agent+5
AI2026-07-02
Claude Fable 5 Returns on July 1, 2026 After a 19-Day Shutdown — U.S. Commerce Department Lifts Export Control Amazon's Jailbreak Research Was the Trigger; Anthropic Applied Mitigations Before Resuming Cursor Restored Fable 5 the Same Evening (Leads CursorBench, Most Expensive Per Task); Devin Fusion + Fable 5 Delivers 41% Cost Reduction vs Fable 5 Alone
Anthropic's Claude Fable 5 and Mythos 5 returned globally on July 1, 2026 after a 19-day shutdown (CNBC / 9to5Mac / VentureBeat / The Hacker News).

Timeline: on June 12 at 5:21 PM ET the U.S. government issued an export-control directive (directly triggered by Amazon researchers reporting a jailbreak) → Anthropic immediately shut down Fable 5 and Mythos 5 worldwide (it had no way to verify user nationality in real time) → June 26: Mythos 5 was cleared for 100+ U.S. institutions → June 30: the Commerce Department lifted controls → July 1: global restoration across Claude Platform / Claude.ai / Claude Code / Claude Cowork / AWS Bedrock / Google Vertex AI / Microsoft Foundry.

Anthropic's mitigations: the Amazon-reported jailpath was addressed, commitments to "proactively detect and address security risks," collaboration with government on future release protocols, and a reporting channel for observed malicious use.

Fable 5 pricing: $10 / M input, $50 / M output (per apidog). Pro / Max / Team / Enterprise subscribers get an extra allocation up to 50% of their weekly usage limits through July 7, then usage-credit billing.

Cursor integration: Cursor officially announced Fable 5 was back at 7:48 PM PT on July 1 on X. "Leads all models on CursorBench — but the most expensive per task," per Cursor (Cursor Community Forum / apidog setup guide).

Devin Fusion: before the shutdown, Devin Fusion + Fable 5 achieved a 41% cost reduction over Fable 5 alone using a parallel two-agent architecture, delivering frontier-level performance at ~35% lower cost (The New Stack).

Enterprise lessons from the 19-day outage: financial services, healthcare, SaaS, and critical-infrastructure firms lost access to production-embedded tools without warning. The phrase "frontier access is now conditional infrastructure" entered the discourse. Cloud-platform diversification (AWS + GCP + Azure) offers no protection when the model itself is restricted. OpenAI reportedly expanded market share during Anthropic's outage through pre-clearance dealings with regulators (MarketScale).

Strategic context: Anthropic released Sonnet 5 on June 30 — the same day export controls lifted. The timing looks intentional: Sonnet 5 is positioned as a fallback to Fable 5, and heading into its IPO, Anthropic now has a clear two-tier stack — frontier Fable 5 for peak capability + mid-tier Sonnet 5 for cost-efficient volume.
Claude Fable 5AnthropicExport Control+4
AI2026-07-01
What Is Cursor for iOS? Control Coding Agents from iPhone and iPad (Composer 2.5, iOS 26+, Public Beta from June 29, 2026)
Cursor (by Anysphere) released its official iOS public beta on June 29, 2026 (official blog / Changelog / App Store / TechCrunch / 9to5Mac).

You can now drive Cursor's coding agents directly from iPhone and iPad (iOS 26.0+ required). Available to all paid-plan subscribers as a public beta, with 75% off Composer 2.5 runs through July 5, 2026.

Main capabilities:
- Launch cloud / background agents or remote-control desktop-running agents (with handoff between the two)
- Voice input for task description, slash commands to guide the agent
- Push notifications + lock-screen Live Activities when work is ready
- Review generated demos, screenshots, logs, and videos; annotate screenshots
- Inspect diffs and merge PRs from the phone
- Model selection across Composer 2.5, GPT (OpenAI), Claude (Anthropic), Gemini (Google), etc.
- MCP integrations — query Datadog logs, summarize Slack activity, etc.

Strategic context: this lands alongside the $60B SpaceX acquisition option on Cursor (referenced in our Grok Build column) and operationalizes the "independent coding agents" thesis of Cursor 2.0 (Oct 2025). Anthropic's Boris Cherny (Claude Code lead) is quoted saying "I do most of my coding on my phone now" (TechCrunch). Joins Claude Code Mobile / Codex Mobile / GitHub Copilot Mobile as the fourth major mobile coding-agent surface.

Caveats: public beta only, the Composer 2.5 promo ends July 5, 2026, no Android or Web yet, requires iOS 26.0+ (older iPhones cannot install), and large-codebase manipulation from mobile alone is still constrained.
CursoriOSMobile+5
AI2026-07-01
Claude Sonnet 5 Deep Dive — Anthropic's June 30, 2026 Release Hits 92.4% on SWE-Bench Verified (+12pt Over Opus 4.6) 1M-Token Context, 88.3% on OSWorld-Verified (Beats the 72.4% Human Expert Baseline), 96.2% on GPQA Diamond, 84.7% on ARC-AGI-2 Introductory $2 / $10 per M Tokens Through August 31, 2026 → Standard $3 / $15, Now the Default in Claude Free / Pro and Claude Code Pro
Anthropic released Claude Sonnet 5 on June 30, 2026 (official release / System Card / TechCrunch / VentureBeat).

The headline: the mid-tier Sonnet just leapfrogged Opus 4.6 by 12 points — 92.4% on SWE-Bench Verified (Opus 4.6 was 80.8%), 88.3% on OSWorld-Verified (15.9 pts ahead of the 72.4% human-expert baseline), 96.2% on GPQA Diamond (over Gemini 3.1 Pro's 94.3%), and 84.7% on ARC-AGI-2 (7.6 pts ahead of Gemini 3.1 Pro's 77.1%). It ships with a 1M-token context window (matching Opus 4.8) and a 128K max output.

Strategic pricing on the eve of Anthropic's IPO: introductory $2 / M input and $10 / M output through August 31, 2026, after which standard pricing becomes $3 / $15 (matching Sonnet 4.6). Note: a new tokenizer maps the same input to about 1.0–1.35× more tokens. It undercuts GPT-5.5, Gemini 3.1 Pro, and Anthropic's own Opus 4.8 on price.

Default-model rollout: now the default in claude.ai Free and Pro, default in Claude Code Pro, and available via API (claude-sonnet-5), AWS Bedrock, Vertex AI, and Managed Agents. Zapier's Daniel Shepard told TechCrunch that "earlier Sonnet versions would stall on multi-step tasks — Sonnet 5 finishes them end-to-end."

Safety: lower misalignment than Sonnet 4.6, cyber safeguards on by default, and a 0.0% exploit-creation rate on Firefox vulnerability tests.

Strategic context: agentic capability is now "table stakes" across foundation-model companies; competition has shifted to cost-efficiency, reliability, and autonomous-task completion. Heading into its IPO, Anthropic is breaking the boundary between its Opus and Sonnet tiers to win the cost / performance contest in high-volume production workloads.
Claude Sonnet 5AnthropicLLM+4
PRAI2026-06-30
[PR] Horiemon AI School Individual Plan / 1-Day Camp at ¥20,000 OFF via Oflight's Affiliate Link Automatic discount (no code entry required; referral code: 5DX4GA)
[ADVERTISEMENT / PR] This article introduces Horiemon AI School Inc. When you sign up via Oflight's special affiliate link, the Horiemon AI School Individual Plan (standard first payment ¥179,080 tax-inc) and the 1-Day Camp (standard ¥220,000 tax-inc) get an automatic ¥20,000 discount → Individual Plan first payment becomes ¥157,080 and the 1-Day Camp becomes ¥198,000 (no code entry required).

About the service: An AI training and consulting business positioned as a "corporate AI adoption partner." Reported figures: 200+ adopting companies, 2,000+ learners, 75%+ AI-deployment cost reduction (per the official LP). The brand is overseen by entrepreneur Takafumi Horie. Both online sessions and in-person sessions in Ichigaya, Tokyo are offered.

Plans:
1. Individual Plan — a monthly subscription AI skill-acquisition program (monthly online training, content access, community participation)
2. 1-Day Camp — two 9am–6pm sessions online (nationwide) or in person (Ichigaya, Tokyo) at ¥220,000 including tax. A distinctive step-up system: Regular → OB Session → Assistant → Instructor → FC (franchise) membership. FC members receive a subsidy of ¥16,750 per month.

Reader benefit: applying through the links below gives you the ¥20,000 OFF automatically (no code entry required on your end). If entering manually, the referral code is 5DX4GA.

Individual Plan signup: https://tinyurl.com/2yk5m5ac
1-Day Camp signup: https://tinyurl.com/27rv7hf3

Conflict-of-interest disclosure: when a signup occurs via a link in this article, Oflight Inc. receives an affiliate fee of ¥30,000 under the affiliate terms. The ¥20,000 reader discount is provided separately by Horiemon AI School and is independent of Oflight's affiliate fee.
ホリエモンAI学校AI研修企業AI導入+3
AI2026-06-29
What Is Open GENAI? Japan's Digital Agency Open-Sources Its Government AI Platform hirokawaguchi/open-genai Brings Full Local Deployment (Keycloak / Ollama / Qdrant / Stable Diffusion / faster-whisper)
Open GENAI (the OSS release of Japan's government generative-AI platform 源内 / GENAI) was published by Japan's Digital Agency on April 24, 2026 on GitHub under MIT license (official release / GENAI Web repo / GENAI AI apps repo).

What's released:
- GENAI Web: an AI interface built on TypeScript / React 19 / Zustand 5 / React Router 7 / AWS CDK / Tailwind CSS, with the Digital Agency design system
- GENAI AI Apps: three development templates — AWS for administrative RAG / Azure for self-hosted LLM / Google Cloud for a legal-system AI referencing current statutes

License: MIT + CC BY 4.0 (commercial use, modification, and redistribution allowed)

Scale: the foundation of a 2026-fiscal-year pilot covering ~180,000 government employees across all ministries, with planned expansion to local governments and private-sector adopters.

Design thesis: REST API-first + an ExApp (external-app integration) microservices model, unifying AWS / Azure / Google Cloud behind a single interface. The goal is to structurally eliminate vendor lock-in and duplicate development across agencies.

vs GenU (AWS Generative AI Use Cases): where GenU leans into AWS managed services (Bedrock Agents / Knowledge Base / MCP), GENAI is REST-API + ExApp-extensible, adding enterprise-grade governance — team management with RBAC (System Admin / Team Admin / User), SAML multi-IdP, KMS CMEK, TTL data-retention policies, multi-layer WAF, Bedrock Inference Profiles. The trade-off: features like video generation, web extraction, and prompt optimization are not built in but delegated to external ExApps.

Validated models: Claude Sonnet 4.6, Amazon Nova Lite. PLaMo 3.0 Prime has also been selected as a trial model.

Caveat: the Digital Agency explicitly states that "permanent maintenance is not guaranteed and the OSS publication may be terminated in the future." Long-term operations are the adopting organization's responsibility.

Fully-local community fork: hirokawaguchi/open-genai (unofficial, experimental, MIT) swaps Cognito → Keycloak (SAML), Bedrock → OpenAI-compatible APIs (Ollama / vLLM / LM Studio), OpenSearch → Qdrant, DynamoDB → SQLite, Transcribe → faster-whisper, Bedrock image → Stable Diffusion — running the full GENAI stack on a single Docker Compose command with zero cloud dependency. Recommended Japanese model: Qwen2.5. Supports macOS Apple Silicon (Metal), Linux + NVIDIA (CUDA), and CPU-only.

Oflight's view: combined with the trends covered in our Local LLM June 2026 Update, Open GENAI is the frontrunner generative-AI platform for Japanese municipalities and public-sector adopters — now via two distinct paths (the official cloud-deployed release plus the hirokawaguchi/open-genai local-deployment fork). The column closes with three direct inquiry funnels for Open GENAI evaluation, custom implementation, and ongoing maintenance.
Open GENAI源内デジタル庁+12
123456789101112131415161718192021222324252627282930313233