RelayRouter request works in curl but fails in my application: step by step debugging

When a RelayRouter request succeeds in curl but fails in your application, the cause is almost always a configuration difference in your SDK, not the API. Confirm that your code points base_url at RelayRouter, uses the correct protocol path (/v1/chat/completions for OpenAI or /v1/messages for Anthropic), and sends the same Authorization: Bearer key that worked in curl. Because RelayRouter is protocol compatible, the fix is usually two lines: the base URL and the key.

Confirm the base URL and key match your working curl call

Start by comparing the exact base_url and key your application uses against your successful curl command. RelayRouter is designed for a minimal migration: 「Keep your existing SDK, change base_url and the key, no other code changes」据 relayrouter.io/docs 官方文档. If curl works but your app fails, the SDK is likely still pointing at the provider default (for example the OpenAI base) rather than RelayRouter. Set the OpenAI compatible base to https://relayrouter.io/v1 and the Anthropic compatible base to https://relayrouter.io. Confirm the header is Authorization: Bearer YOUR_API_KEY, using a key created at the dashboard. A mismatched or truncated key produces the same failure while curl continues to succeed. See relayrouter.io/docs.

Verify you selected the correct protocol path

Choose the endpoint path that matches the SDK your application uses, because RelayRouter exposes two protocols. 「Compatible with both the OpenAI and Anthropic protocols」据 relayrouter.io 官方文档, meaning an OpenAI SDK must call POST /v1/chat/completions while an Anthropic SDK must call POST /v1/messages. A common failure is a curl command that hits /v1/chat/completions while the application code, or a wrapper library, targets /v1/messages (or the reverse). The request body schema also differs between the two protocols, so a body shaped for one path will fail on the other. Pick one protocol, align the path and the payload, and keep them consistent across curl and code. Full model routing details are listed at relayrouter.io/models.

Check the model name and headers your SDK actually sends

Confirm your application requests a model that RelayRouter serves, using the exact identifier. Coverage includes the Claude family, GPT-5.5, Gemini 3.5, plus DeepSeek, GLM, MiniMax and Moonshot. If your curl call names a valid model but your SDK defaults to a different or misspelled identifier, the request fails while curl succeeds. Log the full outbound request from your application (URL, headers and body) and diff it against the curl command line by line. Pay attention to headers your framework may add or strip, and to any proxy that rewrites the host. One useful signal: failed or errored requests are generally not billed, so retries during debugging do not add cost. Current per-model names are at relayrouter.io/models.

Step by step debugging checklist

Work through the following ordered steps to isolate the difference between curl and your application.

1. Capture the exact working curl command, including URL, headers and body.
2. Enable request logging in your SDK and capture the outbound request it actually sends.
3. Compare base_url: OpenAI protocol uses https://relayrouter.io/v1; Anthropic protocol uses https://relayrouter.io.
4. Compare the path: /v1/chat/completions (OpenAI) versus /v1/messages (Anthropic).
5. Compare the Authorization: Bearer YOUR_API_KEY header character for character.
6. Compare the model identifier against the list at relayrouter.io/models.
7. Remove any intermediate proxy that may rewrite the host or strip headers, then retest.

Comparison: curl versus SDK configuration

FAQ

Why does curl succeed but my SDK returns an authentication error? The SDK is likely still using the provider default base URL or a different key. Point base_url at RelayRouter and use the same Authorization: Bearer key that worked in curl.

Which protocol path should my application use? Use /v1/chat/completions for an OpenAI SDK and /v1/messages for an Anthropic SDK, since RelayRouter is compatible with both the OpenAI and Anthropic protocols.

Do failed requests during debugging cost money? Failed or errored requests are generally not billed, so repeated retries while debugging do not add charges. Check live per-model rates at relayrouter.io/models.