v1.0.2

Release v1.0.2

Major feature release focused on full AI Agent compatibility.

Highlights

• Complete tool calling support (round-trip)
• Vision / image input support
• response_format forwarding (JSON mode & structured outputs)
• All binaries rebuilt with latest features

New Features

Full Tool Calling Support

The proxy now fully supports OpenAI-style tool calling:

Assistant messages with tool calls:

{
  "role": "assistant",
  "content": null,
  "tool_calls": [
    {
      "id": "call_xxx",
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": "{\"location\":\"Hanoi\"}"
      }
    }
  ]
}

Tool result messages:

{
  "role": "tool",
  "tool_call_id": "call_xxx",
  "content": "Sunny, 32°C"
}

This enables complete agent tool-calling loops.

Vision / Image Support

ContentPart now accepts image inputs:

{
  "role": "user",
  "content": [
    {"type": "text", "text": "What's in this image?"},
    {"type": "image_url", "image_url": {"url": "https://...", "detail": "high"}}
  ]
}

Images are converted to text representation [image: url] when sent to the xAI backend.

Response Format

response_format is now forwarded:

{
  "response_format": {"type": "json_object"}
}

Supports both simple JSON mode and structured outputs with JSON schema.

Binaries

All platform binaries have been rebuilt:

File	Platform	Size
proxy-windows-amd64.exe	Windows x64	9.2M
proxy-linux-amd64	Linux x64	9.2M
proxy-darwin-amd64	macOS Intel	9.3M
proxy-darwin-arm64	macOS Apple Silicon	8.7M

Upgrade Notes

• Drop-in replacement — no config changes required
• Existing config.json remains compatible
• Agents using tool calling or vision will now work without JSON errors

Testing

Verified flows:

• Tool calling round-trip (assistant → tool → assistant)
• Messages containing images
• response_format passthrough
• All previous features remain functional

Full Changelog

See commit a237bb3.

---

Previous release: v1.0.1 (f44d777)

v1.0.1

Release v1.0.1

Minor release with improved AI Agent compatibility and cross-platform distribution.

Highlights

• message.content now accepts both string and array of content parts (fixes vision/tool-calling agents)
• Pre-built binaries for all major platforms available in bin/
• README cleaned up and updated for current usage

New Features

Flexible Message Content

OpenAIMessage.Content now supports the full OpenAI format:

// Simple string (still works)
{"role": "user", "content": "Hello"}

// Array of content parts (new - required by many AI Agents)
{"role": "user", "content": [{"type": "text", "text": "Hello"}]}

This fixes the error:

json: cannot unmarshal array into Go struct field OpenAIMessage.messages

when connecting AI Agents that use vision or tool-calling formats.

Cross-Platform Binaries

All binaries are now located in bin/:

File	Platform
proxy-windows-amd64.exe	Windows (x64)
proxy-linux-amd64	Linux (x64)
proxy-darwin-amd64	macOS Intel
proxy-darwin-arm64	macOS Apple Silicon

No external dependencies required. ~9MB each.

Changes

• Moved from root-level single binary to organized bin/ folder
• Removed outdated "Environment Variables" section from README (config is now fully config.json driven)
• Updated project structure documentation
• Internal refactor: introduced MessageContent type with custom JSON unmarshaling

Upgrade Notes

• Existing config.json remains fully compatible
• No breaking changes to API or configuration
• Replace old proxy.exe with the appropriate binary from bin/

Testing

All previous flows continue to pass. Additional verification:

• AI Agent sending array content → now succeeds
• All 4 platform binaries execute correctly on target OS

Checksums (example)

Run after download:

sha256sum bin/*

Full Changelog

See commit f44d777 for detailed changes.

---

Previous release: v1.0.0 (cdc9f1e)

v1.0.0

Release v1.0.0

First stable release of xAI OpenAI-Compatible Proxy.

Highlights

• Drop-in replacement for OpenAI chat completions API
• HTTP/2 transport to xAI backend (bypasses Cloudflare 403)
• Streaming and non-streaming responses with proper finish_reason
• Optional API key authentication for multi-user deployments
• Zero-config for single-user (empty api_key)

Features

API Compatibility

• POST /v1/chat/completions — full OpenAI request/response format
• GET /v1/models — returns 5 xAI models
• Streaming SSE with finish_reason: "stop" in final chunk
• Standardized error envelope: {"error": {"message", "type", "code"}}
• CORS headers enabled for browser-based agents

Configuration

• config.json driven (host, port, api_key, headers)
• api_key optional — leave empty to disable auth
• When set, clients must send Authorization: Bearer <key>
• 401 returned for missing or invalid key

Transport

• Go http.Transport{ForceAttemptHTTP2: true} — matches Postman behavior
• 120s timeout, connection pooling
• SSE parser with 1MB buffer for long responses

Usage

# 1. Copy template
cp config.json.example config.json

# 2. Edit config.json
# - Paste Cookie header from console.x.ai (Network tab)
# - Set api_key (optional, empty = no auth)
# - Adjust host/port if needed

# 3. Run
./proxy.exe

# 4. Call with OpenAI SDK
from openai import OpenAI
client = OpenAI(base_url="http://localhost:60443/v1", api_key="not-needed-or-your-proxy-key")
resp = client.chat.completions.create(model="grok-4.3", messages=[{"role":"user","content":"Hi"}])
print(resp.choices[0].message.content)

Authentication

api_key in config	Client header required	Result
"" (empty)	None	200
"sk-xxx"	Authorization: Bearer sk-xxx	200
"sk-xxx"	Missing / wrong key	401

Tested Flows

• Non-stream completion
• Stream with finish_reason
• Multi-turn conversation
• Concurrent requests (3+)
• Long messages (5K+ chars)
• System messages
• All OpenAI optional params passthrough
• CORS preflight
• Method not allowed (405)
• Invalid JSON (400)
• Upstream error wrapping (502)
• Missing/invalid config.json at startup

All 13 scenarios pass.

Limitations

• Token usage always returns 0 (xAI SSE does not emit token counts)
• Some OpenAI params (n>1, certain stop behaviors) may be ignored by backend
• Cookie refresh required when Cloudflare tokens (cf_clearance, __cf_bm) expire (IP-bound)

Files

• proxy.exe — Windows amd64 binary (Go 1.24)
• config.json.example — template with placeholders
• Source included for custom builds

Upgrade Notes

First release — no upgrade path. Fresh install only.

Checksum (example)

SHA256(proxy.exe) = <run sha256sum after download>

Support

Open an issue with:

• config.json (redact Cookie)
• Request/response logs
• Go version and OS

---

Full commit: cdc9f1e — feat: xAI OpenAI-compatible proxy server