AI API Troubleshooting - Unlimited AI | One API for AI Models & Token Plans

By Unlimited AI · Updated September 26, 2026

When an AI coding tool stops responding, find the stage that failed before changing settings. A small reproducible request helps distinguish credentials, routing, model access, limits, and client compatibility.

Capture a minimal failure

Start a fresh session with a short, non-sensitive prompt and the same model. Note the client version, endpoint domain, model ID, time, response status, and error text. Remove API keys, authorization headers, personal data, and private code before sharing logs.

Change one variable at a time. Replacing the endpoint, model, and key together can hide the cause even if a later request works. Keep the last known working configuration so you can compare it with the failure.

Match the symptom to the next check

Symptom	What to check first
Authentication or permission error	Confirm the issuing service, key status, endpoint domain, and account access. Read the provider’s error details.
Model unavailable or not found	Compare the exact model ID with the provider’s current list and your plan’s entitlement.
Route or format error	Check the API protocol and required path. A model name does not identify the correct endpoint.
Rate or quota limit	Inspect whether the error concerns request frequency, concurrent work, balance, or plan allowance.
Context too large	Reduce the conversation and attached files, then check the model’s documented input and output limits.
Chat works; tools fail	Verify tool support and the client’s expected request, result, and streaming formats.

Status codes are clues, not a complete diagnosis. Gateways can wrap upstream errors, so use the response details and provider documentation together.

Separate connectivity from coding quality

First confirm that a basic request completes. Next test a narrow code edit. Finally test the agent features your workflow needs. This sequence shows whether the failure is in the connection, model behavior, or tool orchestration.

If the API responds normally but the proposed code is incorrect, inspect the task instructions, supplied context, and resulting diff. More retries against an unchanged prompt may consume allowance without improving the result. Supply the specific failing check and a smaller scope.

Handle limits and transient failures

When the provider identifies a temporary limit, follow its retry guidance and reduce parallel requests. Avoid a rapid automatic retry loop: it can prolong throttling and create additional usage. An expired allowance or missing model entitlement requires an account or plan correction rather than repeated requests.

For broken or incomplete streams, record whether the issue also occurs with a small request and whether the client reports a protocol error. Use the client’s supported settings; do not disable certificate checks or send credentials to an unverified diagnostic service.

Send support a useful report

Include the tool and version, sanitized configuration, exact error, timestamp with timezone, and a safe reproduction prompt. If available, include a request ID. Explain the last successful step and the single change that preceded the failure.

Unlimited AI customers can use support tickets to confirm delivery details and current model access. Tool-specific setup is available in the coding guides index.