[Discussion] Integrating with GitHub Copilot via `copilot-api` results in an empty `{}` response

[kehao-chen]

Hi there,

First off, I want to say thank you for creating claude-code-router. It's a fantastic project that provides great flexibility in using different AI providers within the Claude Code environment.

I currently have an active GitHub Copilot subscription and was inspired to see if I could route its API through claude-code-router. To achieve this, I'm using another helpful project, ericc-ch/copilot-api, which exposes the GitHub Copilot API as a local, OpenAI-compatible endpoint.

I noticed a similar integration discussion in the copilot-api repository (issue #43), which made me optimistic that this setup could work.

The Problem

However, when I connect claude-code-router to the local endpoint provided by copilot-api, I'm encountering a peculiar issue.

1. copilot-api shows a valid response: Looking at the verbose logs from the copilot-api server, I can clearly see that it's successfully receiving a streaming response from GitHub. The log shows chunks of data containing the content with the generated text.

ℹ Current token count: { input: 3887, output: 27 }                                          11:35:09 AM
⚙ Set max_tokens to: 16384                                                                  11:35:09 AM
⚙ Streaming response                                                                        11:35:10 AM
[11:35:10 AM] ⚙ Streaming chunk: {"data":"{\"choices\":[],\"created\":0,\"id\":\"\",\"prompt_filter_results\":[{\"content_filter_results\":{\"hate\":{\"filtered\":false,\"severity\":\"safe\"},\"self_harm\":{\"filtered\":false,\"severity\":\"safe\"},\"sexual\":{\"filtered\":false,\"severity\":\"safe\"},\"violence\":{\"filtered\":false,\"severity\":\"safe\"}},\"prompt_index\":0}]}"}
--> POST /chat/completions 200 2s
[11:35:10 AM] ⚙ Streaming chunk: {"data":"{\"choices\":[{\"index\":0,\"content_filter_offsets\":{\"check_offset\":18806,\"start_offset\":18806,\"end_offset\":18808},\"content_filter_results\":{\"hate\":{\"filtered\":false,\"severity\":\"safe\"},\"self_harm\":{\"filtered\":false,\"severity\":\"safe\"},\"sexual\":{\"filtered\":false,\"severity\":\"safe\"},\"violence\":{\"filtered\":false,\"severity\":\"safe\"}},\"delta\":{\"content\":\"\",\"role\":\"assistant\"}}],\"created\":1750390510,\"id\":\"chatcmpl-BkMowfAsar7wRdm2kvUMYhXX5WBGi\",\"model\":\"gpt-4.1-2025-04-14\",\"system_fingerprint\":\"fp_07e970ab25\"}"}
[11:35:10 AM] ⚙ Streaming chunk: {"data":"{\"choices\":[{\"index\":0,\"content_filter_offsets\":{\"check_offset\":18806,\"start_offset\":18806,\"end_offset\":18808},\"content_filter_results\":{\"hate\":{\"filtered\":false,\"severity\":\"safe\"},\"self_harm\":{\"filtered\":false,\"severity\":\"safe\"},\"sexual\":{\"filtered\":false,\"severity\":\"safe\"},\"violence\":{\"filtered\":false,\"severity\":\"safe\"}},\"delta\":{\"content\":\"Hi\"}}],\"created\":1750390510,\"id\":\"chatcmpl-BkMowfAsar7wRdm2kvUMYhXX5WBGi\",\"model\":\"gpt-4.1-2025-04-14\",\"system_fingerprint\":\"fp_07e970ab25\"}"}
[11:35:10 AM] ⚙ Streaming chunk: {"data":"{\"choices\":[{\"finish_reason\":\"stop\",\"index\":0,\"content_filter_offsets\":{\"check_offset\":18806,\"start_offset\":18806,\"end_offset\":18808},\"content_filter_results\":{\"hate\":{\"filtered\":false,\"severity\":\"safe\"},\"self_harm\":{\"filtered\":false,\"severity\":\"safe\"},\"sexual\":{\"filtered\":false,\"severity\":\"safe\"},\"violence\":{\"filtered\":false,\"severity\":\"safe\"}},\"delta\":{\"content\":null}}],\"created\":1750390510,\"id\":\"chatcmpl-BkMowfAsar7wRdm2kvUMYhXX5WBGi\",\"usage\":{\"completion_tokens\":3,\"completion_tokens_details\":{\"accepted_prediction_tokens\":0,\"rejected_prediction_tokens\":0},\"prompt_tokens\":11403,\"prompt_tokens_details\":{\"cached_tokens\":9984},\"total_tokens\":11406},\"model\":\"gpt-4.1-2025-04-14\",\"system_fingerprint\":\"fp_07e970ab25\"}"}
⚙ Streaming chunk: {"data":"[DONE]"}                                                        11:35:10 AM
<-- POST /chat/completion

2. claude-code-router shows an empty response: Despite the backend receiving data, the response displayed within the Claude Code UI is consistently an empty object: {}. This happens for various models I've tried to proxy, such as gpt-4o, gpt-4o-mini, etc.

This leads me to believe that the data stream format sent by copilot-api, while OpenAI-compatible, might have a subtle difference that claude-code-router isn't parsing correctly. The data seems to be flowing but gets lost or misinterpreted during processing on the claude-code-router side.

I'm creating this issue to open up a discussion. Has anyone else attempted a similar setup or has any insight into what might be causing this discrepancy? Could it be an issue with how stream terminators ([DONE]) are handled, or a slight variation in the JSON structure of the stream events?

[musistudio]

Interestingly, just a few days ago this discussion also mentioned GitHub Copilot — though it was based on reverse engineering, and the behavior seems different from yours. That might be due to differences in individual implementations.

In the future, I might try to take a closer look at what GitHub Copilot is doing internally to evaluate the feasibility of integrating it. However, since GitHub Copilot has enabled premium request billing, the potential benefit of doing so might be very limited.

[kehao-chen]

Thanks for your reply and for linking that discussion, I hadn't seen it.

You've raised a valid point about the premium request billing. It would certainly be a noticeable limitation for using advanced models. On the bright side, however, the standard GPT-4 level model seems to remain unlimited.

This could still make it a great option to use as a default or "background" model. Thanks again for considering it.

[kehao-chen]

Just wanted to follow up on this thread with a solution I found, in case it helps anyone else.

It turns out the issue was indeed on the ericc-ch/copilot-api side. The implementation in the current release seems to be what was causing the parsing issue within claude-code-router.

After switching to the code from this PR the copilot-api repo #42. I can confirm that it completely resolves the problem.

The integration now works flawlessly, and I'm able to use my GitHub Copilot subscription through claude-code-router, which was my original goal.

Thanks again for the discussion, and I hope this information is useful!

[thesobercoder]

@kehao-chen Can you please share your workflow and config?

[kehao-chen]

@kehao-chen Can you please share your workflow and config?

I ended up not looking into how to use claude-code-router with copilot-api because copilot-api later implemented Anthropic-compatible endpoints. Given that, I didn't think it was worth the extra effort just to get the benefit of claude-code-router's model routing logic.

Step 1. Run copilot-api

npx copilot-api@latest start

Step 2. Set environment variables and start Claude Code

ANTHROPIC_BASE_URL=http://localhost:4141 ANTHROPIC_AUTH_TOKEN=dummy ANTHROPIC_MODEL=claude-sonnet-4 ANTHROPIC_SMALL_FAST_MODEL=gpt-4.1 claude

• Reference document: ericc-ch/copilot-api: Turn GitHub Copilot into OpenAI/Anthropic API compatible server. Usable with Claude Code!