A Single Monthly Credit Pool Now Feeds 11 Chinese LLMs Into Claude Code and Codex
Fragmented AI subscriptions and short-cycle rate limits are a real friction point for developers who switch models per task. A unified credit pool with a single API key removes the overhead of managing multiple accounts and quota windows, though the QPS 2 cap and tool-use restrictions mean it suits solo interactive coding, not batch automation.
Instead of juggling separate subscriptions for GLM, Kimi, DeepSeek, and MiniMax, the AiiOnly Token Plan puts all 11 models into a shared monthly Credits pool. The Standard tier costs 79 yuan/month for 6.32 million Credits, with a QPS limit of 2. Both OpenAI-compatible and Anthropic Messages endpoints are provided, though not every model supports both protocols. A new desktop client, VeryClaw, integrates model access, chat, agents, skills, and scheduled tasks into a single workbench, storing API keys locally.
A walkthrough connects the plan to Claude Code via CC Switch, mapping four role slots — Sonnet, Opus, Fable, Haiku — to DeepSeek-V4-Pro, DeepSeek-V4-Flash, GLM-5.2, and Kimi-K3. The same plan then feeds Codex through the OpenAI Chat Completions protocol, where Kimi-K3 generates a complete HTML/CSS/JS task board in one session. Both tools draw from the same monthly pool, with per-request credit consumption visible in the console.
The plan explicitly forbids use for custom application backends, automation scripts, or batch calls; it is restricted to interactive agents and programming tools. Purchases are non-refundable and cannot be downgraded.
Short-cycle quota windows — 5-hour rolling limits and weekly caps — are a bigger pain point for solo developers than raw model capability, and a monthly pool directly addresses that scheduling friction.
Mapping Claude Code's role slots to different models turns the tool into a model router: backend logic hits a reasoning-heavy model, frontend work hits a multimodal one, all without leaving the session.
The explicit ban on automation scripts and batch calls signals that the pricing is calibrated for interactive, human-paced coding, not for programmatic throughput.
VeryClaw's local API key storage and integrated tool execution point toward a desktop-agent model where the client, not a cloud service, orchestrates model calls and file operations.