AI: Blocked by Cloudflare on OpenAI Compatible

[lutfiandri]

Summary

Conversation in Agent Panel get blocked by Cloudflare, but not with Text Thread on OpenAI Compatible LLM Server

Description

Steps to trigger the problem:

1. Register settings (i wrote below)
2. Set SUMOPOD_API_KEY in environtment
3. Start chat in Agent Panel (got error)
4. Start chat in Text Thread (works, no error)

Expected Behavior:
I expect the Agent Panel can run with OpenAI Compatible from custom server, like the Text Threads.

Actual Behavior:
Chat in Agent Panel always got on error (blocked by cloudflare), but works perfectly fine on Text Threads.

Model Provider Details

• Provider: Sumopod OpenAI Compatible (sumopod.com)
• Model Name: Qwen3 Coder
• Mode: Agent Panel
• Other Details (MCPs, other settings, etc):

Screenshot on Agent Panel

HTML error on Agent Panel

Screenshot on Text Thread (works perfectly fine)

Related Settings

{
  "language_models": {
    "openai_compatible": {
      "Sumopod": {
        "api_url": "https://ai.sumopod.com/v1",
        "available_models": [
          {
            "name": "openrouter/qwen/qwen3-coder",
            "display_name": null,
            "max_tokens": 20000,
            "max_output_tokens": 20000,
            "max_completion_tokens": 20000
          }
        ]
      }
    }
  },

Zed Version and System Specs

Zed: v0.198.3 (Zed)
OS: Linux Wayland fedora 42
Memory: 13.3 GiB
Architecture: x86_64
GPU: AMD Radeon Graphics (RADV REMBRANDT) || radv || Mesa 25.1.4

[rodrigotassinari]

🙋‍♂️ i'm having the exact same issue on Zed 0.199.10 when trying to use anything proxyed by LiteLLM Proxy. Here's one example Ray ID I got: 972d0d9f7a766481

[notpeter]

These is an error from an upstream providers (Sumopod or LiteLLM) that are using Cloudflare -- and their CloudFlare configuration is blocking you.

My best guess for why Text Threads works but not the agent is that the agent will automatically make multiple successive requests which CloudFlare may block, while the Assistant text thread will just be a single request.

If someone has a suggestion for fixes for these providers (e.g. Zed including an additional header on requests, etc) I'm all ears, but as-is this is not something Zed has control over.