codex-shim-byok-models

Codex Shim BYOK Models

Skill by ara.so — Codex Skills collection.

Overview

codex-shim is a local Python API shim that intercepts Codex Desktop's model requests and routes them to:

• Any model in your ~/.factory/settings.json (Factory BYOK)
• Your ChatGPT subscription GPT-5.5
• Custom OpenAI/Anthropic/generic chat completion endpoints

It exposes a local Responses API endpoint that Codex Desktop points to, bypassing Codex's server-side Statsig model allowlist.

Key capabilities:

• Use any OpenAI, Anthropic, DeepSeek, Gemini, OpenRouter, or Z.ai model
• Keep ChatGPT subscription GPT-5.5 alongside BYOK models
• No modification to ~/.codex/config.toml (uses launch-time overrides)
• Translates between Codex Responses API ↔ upstream APIs (OpenAI chat completions, Anthropic messages)

Installation

# Clone repository
git clone https://github.com/0xSero/codex-shim ~/Documents/codex-shim
cd ~/Documents/codex-shim

# Install Python dependencies (requires 3.11+)
python3 -m pip install --user aiohttp pytest

# Symlink commands to PATH
ln -s "$PWD/bin/codex-shim" ~/.local/bin/codex-shim
ln -s "$PWD/bin/codex-app"  ~/.local/bin/codex-app
ln -s "$PWD/bin/codex-model" ~/.local/bin/codex-model

Verify installation:

codex-shim --help

Configuration

Factory Settings File

The shim reads ~/.factory/settings.json by default. Structure:

{
  "customModels": [
    {
      "model": "gpt-5.5",
      "provider": "openai",
      "baseUrl": "https://api.openai.com/v1",
      "apiKey": "${OPENAI_API_KEY}",
      "displayName": "OpenAI GPT-5.5",
      "maxContextLimit": 400000
    },
    {
      "model": "claude-opus-4-7-20251109",
      "provider": "anthropic",
      "baseUrl": "https://api.anthropic.com/v1",
      "apiKey": "${ANTHROPIC_API_KEY}",
      "displayName": "Claude Opus 4.7"
    },
    {
      "model": "deepseek-v4-pro",
      "provider": "anthropic",
      "baseUrl": "https://api.deepseek.com/anthropic",
      "apiKey": "${DEEPSEEK_API_KEY}",
      "displayName": "DeepSeek V4 Pro",
      "noImageSupport": true
    }
  ]
}

Supported providers:

• openai → OpenAI /v1/chat/completions
• generic-chat-completion-api → OpenAI-compatible endpoints
• anthropic → Anthropic /v1/messages

Custom Config Path

codex-shim --settings /path/to/custom-models.json generate
codex-shim --settings /path/to/custom-models.json start

Environment Variables

Store API keys in environment:

export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export DEEPSEEK_API_KEY="..."

Key Commands

Generate Catalog

Reads Factory settings and creates .codex-shim/custom_model_catalog.json:

codex-shim generate

Start/Stop Daemon

# Start shim on 127.0.0.1:8765
codex-shim start

# Check status
codex-shim status

# Stop daemon
codex-shim stop

# Restart
codex-shim restart

List Available Models

# Show all generated slugs and upstream routes
codex-shim list

# Show models currently in Codex Desktop picker
codex-model list

Launch Codex Desktop

# Launch with shim wired in (doesn't modify ~/.codex/config.toml)
codex-app

# Launch and open specific path
codex-app /path/to/project

# Equivalent long form
codex-shim app /path/to/project

Switch Active Model

# List available slugs
codex-model list

# Set default model for next launch
codex-model openai-gpt-5-5

# Relaunch Codex
codex-app

Run Codex CLI Through Shim

codex-shim codex -- chat "explain this code"
codex-shim codex -- edit main.py "add error handling"

Model Routing Architecture

Codex Desktop → /v1/responses → codex-shim (127.0.0.1:8765)
                                   ┃
                    ┏━━━━━━━━━━━━━━╋━━━━━━━━━━━━━━┓
                    ┃              ┃              ┃
         slug "openai-gpt-5-5"    provider       provider
                    ┃            "openai"      "anthropic"
                    ┃              ┃              ┃
         chatgpt.com/backend-api  baseUrl/       baseUrl/
         /codex/responses         chat/          messages
         (Bearer token)           completions    (x-api-key)

The shim:

1. Receives Codex Responses API request
2. Looks up slug in catalog
3. Translates to upstream format (OpenAI chat completions or Anthropic messages)
4. Streams upstream response
5. Translates back to Responses API format

Code Examples

Python: Implementing Custom Provider Translation

# codex_shim/translator.py example pattern

async def translate_to_openai(responses_request):
    """Convert Codex Responses API → OpenAI chat completions."""
    return {
        "model": responses_request["model"],
        "messages": responses_request["messages"],
        "stream": True,
        "temperature": responses_request.get("temperature", 1.0),
        "max_tokens": responses_request.get("max_tokens"),
    }

async def translate_to_anthropic(responses_request):
    """Convert Codex Responses API → Anthropic messages."""
    messages = []
    system = None
    
    for msg in responses_request["messages"]:
        if msg["role"] == "system":
            system = msg["content"]
        else:
            messages.append({
                "role": msg["role"],
                "content": msg["content"]
            })
    
    body = {
        "model": responses_request["model"],
        "messages": messages,
        "stream": True,
        "max_tokens": responses_request.get("max_tokens", 4096),
    }
    
    if system:
        body["system"] = system
    
    return body

Python: Adding Custom Model Programmatically

import json
from pathlib import Path

def add_custom_model(model_config):
    """Add model to Factory settings."""
    settings_path = Path.home() / ".factory" / "settings.json"
    
    if settings_path.exists():
        with open(settings_path) as f:
            settings = json.load(f)
    else:
        settings = {"customModels": []}
    
    settings["customModels"].append(model_config)
    
    with open(settings_path, "w") as f:
        json.dump(settings, f, indent=2)

# Example: Add OpenRouter model
add_custom_model({
    "model": "anthropic/claude-3.5-sonnet",
    "provider": "openai",
    "baseUrl": "https://openrouter.ai/api/v1",
    "apiKey": "${OPENROUTER_API_KEY}",
    "displayName": "Claude 3.5 Sonnet (OpenRouter)",
    "maxContextLimit": 200000
})

Shell: Automated Setup Script

#!/bin/bash
# setup-codex-shim.sh

set -e

CODEX_SHIM_DIR="$HOME/Documents/codex-shim"

# Clone and install
if [ ! -d "$CODEX_SHIM_DIR" ]; then
    git clone https://github.com/0xSero/codex-shim "$CODEX_SHIM_DIR"
fi

cd "$CODEX_SHIM_DIR"
python3 -m pip install --user aiohttp

# Symlink commands
mkdir -p "$HOME/.local/bin"
ln -sf "$CODEX_SHIM_DIR/bin/codex-shim" "$HOME/.local/bin/"
ln -sf "$CODEX_SHIM_DIR/bin/codex-app" "$HOME/.local/bin/"
ln -sf "$CODEX_SHIM_DIR/bin/codex-model" "$HOME/.local/bin/"

# Generate catalog and start
codex-shim generate
codex-shim start

echo "✓ Shim installed and running"
codex-shim status

macOS Picker Patch

Codex Desktop's Statsig config hides models not on a server-side allowlist. Apply this one-time ASAR patch to bypass:

APP=/Applications/Codex.app

# Backup
sudo cp -R "$APP" "$APP.unpatched-$(date +%Y%m%d-%H%M%S)"

# Extract ASAR
cd /tmp && rm -rf codex-asar-patch && mkdir codex-asar-patch && cd codex-asar-patch
npx --yes @electron/asar extract "$APP/Contents/Resources/app.asar" extracted

# Patch picker filter (disables useHiddenModels check)
PATCH_FILE=$(grep -RIl 'useHiddenModels' extracted/webview/assets/model-queries-*.js | head -n1)
sed -i.bak -E 's/let u=c\.useHiddenModels&&o!==`amazonBedrock`,d;/let u=!1,d;/' "$PATCH_FILE"

# Verify exactly one change
diff "$PATCH_FILE.bak" "$PATCH_FILE" && echo "ERROR: No changes made" && exit 1
rm "$PATCH_FILE.bak"

# Repack
npx --yes @electron/asar pack extracted app.asar.new
sudo cp app.asar.new "$APP/Contents/Resources/app.asar"

# Recompute ASAR header hash for Electron integrity check
HEADER_HASH=$(python3 - "$APP/Contents/Resources/app.asar" <<'PY'
import struct, hashlib, sys
with open(sys.argv[1], 'rb') as f:
    data_size, header_size, _, json_size = struct.unpack('<4I', f.read(16))
    header_json = f.read(json_size)
print(hashlib.sha256(header_json).hexdigest())
PY
)

# Update Info.plist
sudo /usr/libexec/PlistBuddy -c \
  "Set :ElectronAsarIntegrity:Resources/app.asar:hash $HEADER_HASH" \
  "$APP/Contents/Info.plist"

# Re-sign (ad-hoc)
sudo codesign --force --deep --sign - "$APP"

echo "✓ Patch applied. Launch Codex Desktop."

Rollback:

sudo rm -rf "$APP"
sudo mv "$APP.unpatched-YYYYMMDD-HHMMSS" "$APP"

ChatGPT GPT-5.5 Passthrough

If ~/.codex/auth.json exists with auth_mode: chatgpt, the shim auto-generates a synthetic slug openai-gpt-5-5 that proxies to:

https://chatgpt.com/backend-api/codex/responses
Authorization: Bearer <access_token from auth.json>

This bypasses Factory and uses your ChatGPT subscription quota.

Disable:

# Remove from catalog after generation
jq 'del(.models[] | select(.slug == "openai-gpt-5-5"))' \
  .codex-shim/custom_model_catalog.json > tmp.json && mv tmp.json .codex-shim/custom_model_catalog.json

Common Patterns

Multi-Provider Setup

{
  "customModels": [
    {
      "model": "gpt-5.5",
      "provider": "openai",
      "baseUrl": "https://api.openai.com/v1",
      "apiKey": "${OPENAI_API_KEY}",
      "displayName": "GPT-5.5"
    },
    {
      "model": "claude-opus-4-7-20251109",
      "provider": "anthropic",
      "baseUrl": "https://api.anthropic.com/v1",
      "apiKey": "${ANTHROPIC_API_KEY}",
      "displayName": "Claude Opus 4.7"
    },
    {
      "model": "gemini-2.0-flash-exp",
      "provider": "openai",
      "baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai",
      "apiKey": "${GOOGLE_API_KEY}",
      "displayName": "Gemini 2.0 Flash"
    }
  ]
}

OpenRouter Aggregation

{
  "model": "meta-llama/llama-3.3-70b-instruct",
  "provider": "openai",
  "baseUrl": "https://openrouter.ai/api/v1",
  "apiKey": "${OPENROUTER_API_KEY}",
  "displayName": "Llama 3.3 70B (OpenRouter)"
}

DeepSeek with Anthropic-Style Thinking

{
  "model": "deepseek-v4-pro",
  "provider": "anthropic",
  "baseUrl": "https://api.deepseek.com/anthropic",
  "apiKey": "${DEEPSEEK_API_KEY}",
  "displayName": "DeepSeek V4 Pro",
  "noImageSupport": true
}

The shim translates extended thinking blocks from thinking items to reasoning.encrypted_content.

Troubleshooting

Models Don't Appear in Picker

Cause: Statsig allowlist still active.

Solution: Apply macOS picker patch (see section above).

"Connection refused" When Launching Codex

Cause: Shim daemon not running.

codex-shim status  # Check if running
codex-shim start   # Start if stopped

401 Unauthorized Errors

Cause: API key not found or invalid.

Check environment variables:

echo $OPENAI_API_KEY
echo $ANTHROPIC_API_KEY

Verify settings file:

cat ~/.factory/settings.json | jq '.customModels[].apiKey'

Test upstream directly:

curl https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY"

ChatGPT Passthrough Not Working

Verify auth.json exists:

cat ~/.codex/auth.json | jq '.auth_mode'
# Should output: "chatgpt"

Check access token validity:

TOKEN=$(jq -r '.access_token' ~/.codex/auth.json)
curl https://chatgpt.com/backend-api/me \
  -H "Authorization: Bearer $TOKEN"

Shim Logs

# Tail daemon logs
tail -f ~/.codex-shim/codex-shim.log

# Check last 50 lines
tail -n 50 ~/.codex-shim/codex-shim.log

Port Already in Use

# Change default port (8765)
codex-shim --port 9876 start
codex-shim --port 9876 app

Extended Thinking Not Appearing

For Anthropic-shaped providers (Claude, DeepSeek), thinking blocks appear as reasoning.encrypted_content items in Codex UI.

Verify provider is set to anthropic:

jq '.customModels[] | select(.model=="deepseek-v4-pro") | .provider' \
  ~/.factory/settings.json

Should return "anthropic", not "openai".

Testing

Run test suite:

cd ~/Documents/codex-shim
python3 -m pytest tests/ -v

Test specific translation:

# tests/test_translation.py
import pytest
from codex_shim.translator import translate_to_anthropic

@pytest.mark.asyncio
async def test_system_message_extraction():
    request = {
        "model": "claude-opus-4-7",
        "messages": [
            {"role": "system", "content": "You are helpful"},
            {"role": "user", "content": "Hello"}
        ]
    }
    
    result = await translate_to_anthropic(request)
    assert result["system"] == "You are helpful"
    assert len(result["messages"]) == 1

MCP Tool Forwarding

Codex Desktop forwards three generic MCP tools to all models (built-in and shim-routed):

• list_mcp_resources
• list_mcp_resource_templates
• read_mcp_resource

These are available in the function calling schema for every routed model. The model calls list_mcp_resources to discover available resources.

Note: Codex Desktop does not flatten individual MCP server tools. That's a Codex client behavior, not a shim limitation.

---

Official Repository: https://github.com/0xSero/codex-shim
License: MIT
Requires: Python 3.11+, aiohttp