Is This Claude API Relay Genuine?

When Claude thinks through a problem, Anthropic's server stamps a cryptographic "fingerprint" (like an official Anthropic seal on a package). This seal can only be created with Anthropic's own private key — relays cannot forge it. TokenAPI Scan checks for this seal — if the relay's response has no seal or a wrong one, it's not real Claude and may be running a cheaper model instead.

The thinking signature is the signature field returned in Claude's response when extended thinking is enabled — a server-side encrypted artifact 500-2000 characters long. This signature is generated by Anthropic's servers with cryptographic verification — relays theoretically cannot forge it. TokenAPI Scan assigns it 25% weight as the core detection item. Any relay that cannot return this signature (or has an abnormal length) is flagged as non-genuine Claude.

Three-step foolproof judgment: ① Run TokenAPI Scan — score above 85 is basically usable, 100 usually means 1:1 official passthrough; ② Check for red "severe" markers — if present, it may be swapping models, switch to another provider; ③ Check the provider's leaderboard history to see what others found — 10+ tests with consistently high scores is more trustworthy. Cheap relays are often watered down — score matters more than price.

Focus on 4 things: ① Run TokenAPI Scan full mode — check if thinking signature scores 100; ② Check the usage field for non-claude_* origins (e.g. usage_source: anthropic in OpenAI responses indicates relay wrapping); ③ Check stream/non-stream input_tokens / output_tokens difference on the same prompt — should be ≤ 5%; ④ Check if message IDs follow msg_ / toolu_ / srvtoolu_ prefix conventions — UUID-like IDs are a common relay forgery pattern.

Some relays actually use Amazon's "wrapped Claude" backend (products called Kiro or Amazon Q). Amazon's version strips out Claude's thinking "fingerprint" — because they use a simplified interface. TokenAPI Scan checks — no Anthropic fingerprint seal = immediately flagged not real Claude. Simply put, TokenAPI Scan uses a cryptographic seal that relays cannot forge as its truth mirror — impostors are exposed in 5 seconds.

Kiro and Amazon Q are Amazon's Claude proxy gateways that use AWS Bedrock's simplified interface — their responses do not include the thinking signature (signature generation requires Anthropic's private key, which Bedrock does not implement). TokenAPI Scan scores 0 on the thinking_signature detector, zeroing out the 25% weight and locking the overall verdict to marginal or failed. The same applies to other Bedrock-routing relays.

No. Your API key is only used temporarily during the detection (tens of seconds), and is cleared from memory immediately after — it is never written to any file, log, or database. TokenAPI Scan's servers will never find your key. The report page shows a masked version (like sk-y7xU••••••0h). Our code is open source — if concerned, you can audit it or run it locally on your own machine.

No. The API key exists only in the asyncio.Job object — once the job finishes (success / failure / timeout), it is immediately released from memory. It is not written to report JSON, access logs, or persisted anywhere. The key in the report is masked via mask_api_key() (first 4 + last 4 characters preserved). Full code is MIT-licensed , self-hostable — Docker deployment in 5 minutes.