Claude API relay detection Crypto-grade check

Verify whether a Claude API relay truly passes through

pangolin uses thinking signature, protocol fields, and capability probes to tell native Claude from wrappers or cut-down backends.

Thinking Signature 20 checks Long context / PDF Keys not stored

This page emphasizes server-side signature verification. Real requests are billed to the key’s upstream account.

Claude highlight: closer to identity-grade verification

Claude detection is not “does it look like Claude?” — it checks whether server signatures and Anthropic protocol evidence actually exist.

Thinking signatureAnthropic server-side signature — relays cannot forge it in theory. Highest-confidence signal on this page.
Capabilities & anti-injectionTool Use, PDF, long context, streaming consistency, plus system-conflict / minimal-input audits / official same-prompt compare.
Safety boundaryAPI keys are used only for this job — never written to reports, logs, or disk.

For crypto-grade identity, start with Claude; for protocol compatibility, also check OpenAI / Gemini.

Relay base URL

Enter the relay root URL. pangolin will call its /v1/messages endpoint.

API key

The key is used only for this job and is never written to reports or durable storage.

Target model

Pick a preset or type the custom model ID your relay uses.

Scan depth

Long-context authenticity probe Extra 32k → 100k → 200k needle-in-haystack requests. Extra upstream cost, typically about $0.05–$1.00.

Extreme context probe Steps up toward the model’s advertised window (up to ~950k). Cost may reach $0.05–$8.

Checks that will run

Updates with Scan depth above. Full mode covers all applicable checks. Long-context probes need an extra opt-in.

Full · all applicable checks (recommended) · 20

Model authentication

Examine the model's self-reported identity, brand residue, and common signs of camouflage.

Think supportive

Verify whether server-side exclusive fields such as Claude thinking / signature are transparently transmitted.

Accurate command reproduction (anti-casing)

The model is asked to reproduce a fixed mark verbatim, checking how deterministic the instructions are followed.

System conflict testing

Compare user only to explicit system+user, checking whether fixed words are overridden by hidden directives.

Minimalist Input Token Audit

Use minimalist requests to observe whether input_tokens is abnormally high and identify implicit System injection.

Extreme reasoning ability (number theory counting)

Use canned reasoning questions to check whether verifiable answers are correct.

Channel detection

Identifies the hosting channel from the response ID prefix.

Claude Tokenizer verification

Count input_tokens with fixed hints compared to calibrated model baseline.

behavioral style consistency

Compare Claude series characteristics through behavioral preferences and answer patterns.

knowledge accuracy

Use Anthropic / Claude general knowledge questions to cross-validate whether the backend is as claimed as Claude.

Model declaration consistency

Comparing request model, response model and identity stability in multiple rounds of returns.

Message structure specification

Check message IDs, SSE event sequences, and content blocks for compliance with Anthropic specifications.

Tool Use/Structured Output

Check whether the tool_use block, toolu_ ID, function name and parameter schema are standardized.

streaming consistency

Compare stream and non-stream text, token, and end reason for consistency.

Token usage consistency

Check that the input / output / cache token fields are complete and authentic.

Official comparison of the same title

The same question compares input consumption and response behavior between channels and official (or baseline).

long context authenticity

Use needle-in-haystack to verify that long context windows are honored.

Multimodal/PDF recognition

Submit a PDF probe to confirm that the document understanding link has not been stripped by the transfer station.

Hide injection detection

Randomized variants cross-check for abnormal signals such as nonce leaks, fixed replies, and suspicious cache reuse.

Security rejection and policy stability

Check if the security policy is stable using denial-of-answer scenarios and rewriting/encoding/progressive variants.