Status and limits
What the platform is, what it does when a dependency is down, and the things it will not do at all.
Current state
| Public API | accepting traffic |
| Catalogue | 214 endpoints, version 1.0.0 |
| Settlement | x402 on X Layer, stablecoin |
| Published response examples | 97 captured from real calls; the rest explain why they have no fixed example |
When something upstream is down
Endpoints degrade by refusing, not by guessing. Each returns a specific code naming what was unavailable.
| Dependency | Affects | What happens |
|---|---|---|
| OKX public market API | market, analytics | Endpoints return a dependency error naming the source. They never fall back to a cached price presented as current. |
| X Layer RPC | security, xlayer | Screening returns RPC_UNAVAILABLE rather than a clean bill of health. An absent answer is not a passing one. |
| Local AI model | ai | All 14 AI endpoints return AI_PROVIDER_NOT_CONFIGURED. They never substitute generated placeholder text for a real answer. |
| Vision model | OCR-style AI endpoints | A scanned page with no text layer is refused rather than handed to a text-only model, which would produce confident prose about a document it never saw. |
| LibreOffice | DOCX ↔ PDF conversion | Returns PROVIDER_NOT_CONFIGURED with install instructions rather than failing obscurely. |
What this platform will not do
- Trade, withdraw, transfer or sign anything. It holds no private key. Every signing and sending RPC method is refused at the provider layer, and a test asserts the provider class exposes no such method.
- Execute what you upload. Documents are parsed, never run. Macros are detected and reported. Executables are refused on their actual bytes, whatever the filename says.
- Fetch what it should not. Outbound requests are blocked from localhost, private ranges, link-local addresses and cloud metadata endpoints, with DNS re-resolved after every redirect. robots.txt is enforced, not merely noted.
- Invent a finding. Security screening reports the on-chain facts it found. Competitor and evidence summaries quote your own source material, and a quotation that cannot be located there is dropped rather than published.
- Let a model do arithmetic. On AI endpoints the model writes prose; every total is recomputed with decimal arithmetic and both figures are returned, along with the difference.
Per-endpoint limits
Request size, timeout and concurrency are set per endpoint and published on each one's page. They are enforced before the expensive work starts, not after.