Response format incompatible with Clawdbot's OpenAI provider

[gebeer]

Context

Attempted to use this proxy with Clawdbot as documented in Clawdbot's docs (docs/providers/claude-max-api-proxy.md). The README here suggests Clawdbot users should use native OAuth instead, but Clawdbot's docs provide a proxy config example.

Configuration Used

{
  "env": {
    "OPENAI_API_KEY": "not-needed",
    "OPENAI_BASE_URL": "http://localhost:3456/v1"
  },
  "models": {
    "providers": {
      "openai": {
        "baseUrl": "http://localhost:3456/v1",
        "api": "openai-completions",
        "models": [
          { "id": "claude-opus-4", "name": "Claude Opus 4", "reasoning": true }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "openai/claude-opus-4" }
    }
  }
}

Observed Behavior

1. Proxy health check: ✅ OK
2. /v1/models endpoint: ✅ Returns expected model list
3. Chat completion request: ❌ Response received but content is [object Object]

Clawdbot replied with:

"I notice your messages appear to be showing [object Object] which typically indicates a serialization issue..."

Expected Behavior

Response content should be a properly serialized string, not a JS object reference.

Likely Cause

The cli-to-openai.ts adapter may be returning an object where a string is expected in the response content field, or the SSE stream chunks aren't being serialized properly.

Suggestion

Either:

1. Add tested Clawdbot integration config to this README (and verify it works)
2. Or clarify in README that Clawdbot users should use native OAuth (and coordinate with Clawdbot docs to remove/update the proxy guide)

[kevinfealey]

@gebeer I ran into the same issue. I just fixed it in PR #11 if you want to try it - it's working for me. I also updated the README with the config that's working. Hope it helps!

[gebeer]

@gebeer I ran into the same issue. I just fixed it in PR #11 if you want to try it - it's working for me. I also updated the README with the config that's working. Hope it helps!

We moved away from proxying and are going through the claude-cli provider instead. But thanks for the heads-up!

[wende]

@gebeer @kevinfealey
This repository generally has a couple of flaws.

1. Incorrect OpenAI format interpretation
2. Crashes on rate limits instead of gracefully failing
3. Passes OpenClaw tools as a system prompt, even though claude has it's own toolset that conflicts with that

I've made a fork that solves these. Where nr 3 is solved by --append-system-prompt instead of replacing it, as well as adding another system prompt that translates all tools (exec -> Bash etc) so claude code can interpret skills and whatnot properly.
So if you're still interested it's here

https://github.com/wende/claude-max-api-proxy

P.S fellow openclawers, read all installed skills, i got caught my openclaw making requests to an IP number encoded with base64 in a bash command

[orion-xlabs]

@wende

@gebeer @kevinfealey This repository generally has a couple of flaws.

1. Incorrect OpenAI format interpretation
2. Crashes on rate limits instead of gracefully failing
3. Passes OpenClaw tools as a system prompt, even though claude has it's own toolset that conflicts with that

I've made a fork that solves these. Where nr 3 is solved by --append-system-prompt instead of replacing it, as well as adding another system prompt that translates all tools (exec -> Bash etc) so claude code can interpret skills and whatnot properly. So if you're still interested it's here

https://github.com/wende/claude-max-api-proxy

P.S fellow openclawers, read all installed skills, i got caught my openclaw making requests to an IP number encoded with base64 in a bash command

Good stuff. i was following the process on the openclaw documentation and i'm suprised they documented a version that wasn't working fully at least.
Concerning the installed skills, are you referring to any skills in particular that you installed?

[ksylvan]

@wende Are you contributing your fixes to @atalovesyou repo?

[wende]

@ksylvan the repo seems to be dead. 2 commits and not a single contact from the developer

@orion-xlabs it was a YouTube transcribe skill. With youtube-jop if I remember correctly. But I'm pretty sure it was just copied from a fully functioning one with just added malicious line

[sblanchard]

fixed in last pr