litellm.AuthenticationError: invalid x-api-key when using Anthropic

[hanbit-beaubrain]

Git provider

Github Cloud

System Info

Describe the bug
I am currently using pr_agent with Claude (Anthropic) as my LLM. Starting about 2 days ago, I began encountering a litellm.AuthenticationError stating invalid x-api-key during LLM inference.

To Reproduce

Run pr_agent configured to use Anthropic (Claude).
The error occurs during the chat_completion phase in litellm_ai_handler.py.
Expected behavior
Successful LLM inference using the Anthropic API without authentication errors.

Logs / Error Message
Here is the error log I'm getting:

{"text": "Error during LLM inference: litellm.AuthenticationError: AnthropicException - {"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"},"request_id":"req_011CZRJKw2HYvZXX8mSK2LHr"}\n", "record": {"elapsed": {"repr": "0:00:25.615212", "seconds": 25.615212}, "exception": null, "extra": {}, "file": {"name": "litellm_ai_handler.py", "path": "/app/pr_agent/algo/ai_handlers/litellm_ai_handler.py"}, "function": "chat_completion", "level": {"icon": "⚠️", "name": "WARNING", "no": 30}, "line": 418, "message": "Error during LLM inference: litellm.AuthenticationError: AnthropicException - {"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"},"request_id":"req_011CZRJKw2HYvZXX8mSK2LHr"}", "module": "litellm_ai_handler", "name": "pr_agent.algo.ai_handlers.litellm_ai_handler", "process": {"id": 7, "name": "MainProcess"}, "thread": {"id": 140704083827584, "name": "MainThread"}, "time": {"repr": "2026-03-26 07:30:07.944313+00:00", "timestamp": 1774510207.944313}}}

Additional context

LLM Provider: Anthropic (Claude)
This issue started happening ~2 days ago. It was working fine previously.

Bug details

{"text": "Error during LLM inference: litellm.AuthenticationError: AnthropicException - {"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"},"request_id":"req_011CZRJKw2HYvZXX8mSK2LHr"}\n", "record": {"elapsed": {"repr": "0:00:25.615212", "seconds": 25.615212}, "exception": null, "extra": {}, "file": {"name": "litellm_ai_handler.py", "path": "/app/pr_agent/algo/ai_handlers/litellm_ai_handler.py"}, "function": "chat_completion", "level": {"icon": "⚠️", "name": "WARNING", "no": 30}, "line": 418, "message": "Error during LLM inference: litellm.AuthenticationError: AnthropicException - {"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"},"request_id":"req_011CZRJKw2HYvZXX8mSK2LHr"}", "module": "litellm_ai_handler", "name": "pr_agent.algo.ai_handlers.litellm_ai_handler", "process": {"id": 7, "name": "MainProcess"}, "thread": {"id": 140704083827584, "name": "MainThread"}, "time": {"repr": "2026-03-26 07:30:07.944313+00:00", "timestamp": 1774510207.944313}}}

[eslopezag]

The same is happening to me when using Gemini. Setting a dummy OPENAI_KEY env var is a current workaround, and an AI-powered analysis of the bug gives me this likely root cause:

PR #2278 added this line in litellm_ai_handler.py (line 409):

kwargs["api_key"] = litellm.api_key

This unconditionally passes litellm.api_key as the api_key kwarg to every acompletion() call.

The problem:

1. When OPENAI_KEY is not set and OPENAI_API_KEY is not in the environment, litellm.api_key remains None.
2. LiteLLM normally resolves the correct API key per-provider internally (e.g. reading GEMINI_API_KEY for Gemini models). But when api_key=None is explicitly passed as a kwarg, it overrides LiteLLM's own provider-specific key resolution, causing the failure.
3. Setting something like OPENAI_KEY: sk_dummy triggers the __init__ fallback that sets litellm.api_key = "dummy_key". LiteLLM then sees a non-None string that doesn't match any provider format and correctly falls through to mode-specific key resolution — which is why the workaround works.

Suggested Fix:

Make the api_key kwarg conditional so it doesn't override LiteLLM's per-provider key resolution:

-            kwargs["api_key"] = litellm.api_key
+            if litellm.api_key and litellm.api_key != "dummy_key":
+                kwargs["api_key"] = litellm.api_key

Or scope it to only Ollama models, which was the intent of the original PR:

-            kwargs["api_key"] = litellm.api_key
+            if get_settings().get("OLLAMA.API_KEY", None):
+                kwargs["api_key"] = litellm.api_key

Again, I clarify that the analysis and suggestion above come from an AI agent, and I don't purport to have experience with this codebase, so take it with a grain of salt.

[yanukadeneth99]

Hey @hanbit-beaubrain, could you check whether this issue is fixed on the latest v0.34?