Turn AI Dependency Into a Portfolio — Claude / Codex / Jibun Inc. on a Balance Sheet
Where This Post Sits
In the earlier post "Claude Code vs OpenAI Codex Desktop vs Your Life Hub — A 3-Layer Design for Bundling AI" I mapped how three tools operate on different layers and therefore don't compete.
This post is the sequel: "why bundling is the rational choice, told through accounting." Pretend you keep a mental balance sheet (BS) as a personal CEO, and decompose AI dependency into "assets" and "liabilities."
Shared Fact Base (2026-04-17 Event)
On 2026-04-17, OpenAI Codex Desktop shipped Computer Use + 20+ plugins + MCP servers. Initial headlines claimed Codex had reached plugin parity with Claude, but a cross-source check gives the real numbers:
- Claude Code: 423 plugins / 2,849 skills / 177 agents (43 marketplaces / 834 total)
- OpenAI Codex: 20+ plugins (no self-serve publish yet)
→ Claude leads plugins by ~20×. What Codex caught up on is Computer Use only.
Keep these numbers in mind for the next section.
The Personal-CEO BS Principle (Jibun Inc. Principle #7)
Principle #7 of Jibun Inc.'s 9 principles is "Asset/Liability Balance Sheet." Clean definition:
Things that increase future revenue / time / autonomy = Asset
Things that decrease future revenue / time / autonomy = Liability
Every daily choice thickens one side of the BS — employee or indie dev. AI tool choice is no exception.
Decompose AI Tools on a BS
Single-vendor lock-in, viewed through accounting, is a short-term liability with extremely low liquidity. Why:
- Switching cost: memory / skill / plugin / workflow are vendor-specific → rebuild cost on exit
- Price risk: zero veto on price hikes (recall Anthropic $20 → $100/seat)
- Availability risk: outages stop your work (no redundant staff for a solo CEO)
- Roadmap risk: vendor strategy pivots can obsolete your workflow
Concrete mapping:
| AI Tool | Short-Term Liability (-) | Hidden Asset (+) | Net |
|---|---|---|---|
| Claude Code single-dep | Anthropic total dependency / 423 plugins become useless if Anthropic pulls them | World's largest plugin ecosystem | Liability dominant |
| OpenAI Codex single-dep | ChatGPT ecosystem dependency / Computer Use 1-month head start | 3M weekly DAU info stream | Liability dominant |
| Cursor single-dep | Anysphere dep + IDE lock-in | Seamless in-IDE UX | Liability dominant |
| Jibun Inc. (ai-hub distributed) | Initial build cost | Command post bundling Claude + OpenAI + Gemini + fallback / auto-failover on outage | Asset dominant |
"Claude is strong at 423 plugins" is a fact. It does not justify single-vendor dependency. If anything, 423 > 20 amplifies the other face: higher exit cost = tighter lock-in that narrows future options.
Sidebar: credit-metered pricing = monthly consumptive liability
On 2026-05-04, Notion Custom Agents switches to $10 / 1,000 credits (triangulation lands a conservative $0.11–$0.33 per agent run). Credit-based metered pricing has an extra liability characteristic in accounting terms:
- Monthly reset: unused credits evaporate → not a deferrable asset (the logic "I prepaid so this month is cheaper" does not apply)
- Usage pressure: end-of-month "I need to burn remaining credits" is a perverse incentive where saving budget and using AI become zero-sum
- Personal CEO's time capital: ~100 seconds/day spent checking credit balance × 365 = ~10 hours/year of "balance-check labor"
Unlike flat-rate subscription, credit-metered pricing belongs on the BS as "short-term, non-deferrable, usage-pressure liability."
By contrast, Jibun Inc. (Supabase + Flutter Web) has zero credit concept → structurally does not carry this liability.
Stage 2: mandatory consumption commits (Anthropic Enterprise, 2025-11)
Around the same time, Anthropic Enterprise moved from $200 flat to $20/seat + usage (verified across Anthropic's own pricing + The Register + 4 more = 6 sources). The structural core is the mandatory consumption commit:
- Pre-pay Anthropic-estimated monthly token volume → if actual usage falls below, you're still billed the full amount
- Prior 10–15% volume discounts are gone
- Redress Compliance estimates heavy users land at 2–3× cost increase
On the BS, commit pricing is a "monthly mandatory liability" — arguably worse than credit-metered. With credits you can at least "burn them by month-end for parity"; with commits, the under-consumption is simply lost — zero deferral + no penalty for using less = fully one-sided.
Looking at Notion (credit-metered / monthly reset) + Anthropic (commit-based / monthly mandatory) as a two-stage rocket, the pattern is clear: vendors stabilize their revenue via metered + commit pricing; users absorb the unpredictability.
The counter is simple: don't bet on one AI vendor. Bundle via ai-hub so any single pricing update dilutes rather than detonates your BS.
How Jibun Inc. Implements This
ai-hub: Sit On the Side That Doesn't Pick
// Pseudocode: Flutter Web call
final response = await supabase.functions.invoke('ai-hub', body: {
'action': 'provider.chat_auto',
'messages': [...],
'intent': 'design_decision', // → routes to Claude
});
Edge Function routes by intent:
| intent | 1st choice | fallback 1 | fallback 2 |
|---|---|---|---|
design_decision (long memory) |
Claude Sonnet | GPT-4.1 | Gemini 2.5 Pro |
computer_use_required |
Codex (plugin) | Claude Computer Use | — |
cost_sensitive_bulk |
Haiku | Gemini Flash | DeepSeek |
fallback_any |
Gemini Flash | Haiku | Mistral Small |
cost-hub: 4-Stage Circuit Breaker
- green: normal (all models)
- yellow: per-session cost > $0.50 → sonnet → haiku downgrade
- orange: $1.50 → force downgrade to flash / deepseek-v3
- red: $3.00 → stop calls + alert
→ Single-vendor price hikes can't whip you around. If Anthropic raises prices, traffic shifts to Gemini / DeepSeek.
Let Supabase Own the History
The tool layer stays vendor-agnostic; Jibun Inc.'s PostgreSQL holds history:
- Sleep / expenses / learning / health / KPIs → Supabase tables
- Claude sessions are volatile / Codex plugin invocations are one-shot
- → "Switch AIs, keep your past self's data." Fixed asset that doesn't depreciate.
The Accounting View
Personal-CEO BS (capital = time / attention / autonomy), simple T-account:
【Assets (+)】 【Liabilities (-)】
- 6-dept KPI history (Supabase) - Single-vendor AI dep
- ai-hub routing asset - Manual credit-balance watching
- Hand-rolled hooks / workflows - Silent pauses (billing flip)
- Cross-searchable memory/ - English-first UX cognitive cost
【Net worth】
= Assets − Liabilities
= ai-hub × 6-dept integration × Japanese-native UX
Cancel a single-tool subscription → Jibun Inc.'s net worth doesn't drop. That's the core of the "bundle" design.
Conclusion
- "Which AI to pick" = a decision about the composition of your short-term liabilities
- "Do I have a hub that bundles AIs" = a decision about whether you own a fixed asset
- The rational play for a personal CEO is the latter: thicken the asset side of your own BS
Claude's 423 plugins are amazing. Codex's Computer Use is amazing. But there's zero reason to bet on one. Bundle them, and both strengths land on your asset side.
References
- Companion piece (3-layer map): Claude Code vs OpenAI Codex Desktop vs Your Life Hub
- 9 principles: https://my-web-app-b67f4.web.app/philosophy
- 21-competitor comparison: https://my-web-app-b67f4.web.app/comparison
- Live: https://my-web-app-b67f4.web.app/
Top comments (0)