Codex CLI Integration Guide

Codex CLI is OpenAI’s open-source agentic coding assistant that runs in your terminal. It can read files, write code, run commands, and iterate on its work using the Responses API.

Note

Requires Responses API - Codex CLI uses the /v1/responses endpoint which Infercom supports with MiniMax-M2.7.

Caution

Known limitation: Tool-using (agentic) sessions via the Responses API are currently affected by a server-side streaming issue and may return empty responses for tool calls. For reliable agentic coding today, use OpenCode, Cline, or Aider, which run over the Chat Completions API. This page will be updated when resolved.

Prerequisites

• Node.js 18 or later
• npm or yarn
• Infercom API key

Installation

npm install -g @openai/codex

Verify installation:

codex --version

Caution

macOS users: If macOS blocks the binary as “malware”, go to System Preferences > Security & Privacy and click “Allow Anyway”, or run:

xattr -d com.apple.quarantine $(which codex)

Configuration

Codex CLI supports custom providers via a TOML configuration file.

Step 1: Set Environment Variable

export INFERCOM_API_KEY="your-infercom-api-key"

Tip

Add this to your shell profile (~/.bashrc, ~/.zshrc) for persistence.

Step 2: Create Config File

Create ~/.codex/config.toml:

# Define Infercom as a provider
[model_providers.infercom]
name = "Infercom (EU Sovereign)"
base_url = "https://api.infercom.ai/v1"
env_key = "INFERCOM_API_KEY"
wire_api = "responses"

# Default profile using MiniMax-M2.7
model_provider = "infercom"
model = "MiniMax-M2.7"

Tip

For complex tasks requiring deeper reasoning, Codex CLI supports profiles with different models. You can use a frontier model (Claude, GPT, Gemini) for planning and MiniMax-M2.7 for execution. See Codex documentation for profile configuration.

Step 3: Verify Setup

Run Codex in any project directory:

codex

You should see Codex start with:

model: MiniMax-M2.7
provider: infercom

Model

Use MiniMax-M2.7 - optimized for agentic coding with 192K context and built-in reasoning.

Usage

Interactive Mode

Start Codex in your project directory:

cd your-project
codex

Type your request and Codex will:

• Read relevant files
• Write or edit code
• Run terminal commands
• Iterate until the task is complete

Example Tasks

• “Add error handling to the login function”
• “Write unit tests for the User class”
• “Refactor this file to use async/await”
• “Find and fix the bug causing the test to fail”

Non-Interactive Mode

For scripted or one-shot use, use the exec subcommand:

codex exec "Add input validation to auth.py"

Configuration Options

Full ~/.codex/config.toml reference:

# Default model to use
model = "MiniMax-M2.7"

# Provider to use
model_provider = "infercom"

# Approval mode: "suggest" (default), "auto-edit", or "full-auto"
approval_mode = "suggest"

# Sandbox mode for command execution
sandbox = "docker"  # or "none" to disable

[model_providers.infercom]
name = "Infercom (EU Sovereign)"
base_url = "https://api.infercom.ai/v1"
env_key = "INFERCOM_API_KEY"
wire_api = "responses"

Approval Modes

Mode	Behavior
suggest	Shows diffs and asks before applying (default)
auto-edit	Automatically applies file changes, asks for commands
full-auto	Automatically applies all changes

Troubleshooting

Connection Errors

Verify your configuration:

# Check API key is set
echo $INFERCOM_API_KEY

# Test API directly
curl -s https://api.infercom.ai/v1/responses \
  -H "Authorization: Bearer $INFERCOM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"MiniMax-M2.7","input":"Hello"}' | jq .status

Expected output: "completed"

Model Not Found

Ensure the model name is exact (case-sensitive): MiniMax-M2.7

Slow Responses

MiniMax-M2.7 runs at 400+ tokens/sec. If responses seem slow:

1. Check your network connection
2. Large context (many files) increases processing time
3. First request may be slower due to model loading

Why Codex CLI with Infercom?

Feature	Benefit
EU Sovereign	Data processed in Germany, GDPR compliant
Responses API	Native support for Codex’s agentic architecture
Fast inference	400+ tokens/sec with MiniMax-M2.7
No vendor lock-in	Open-source tool, standard API

Next Steps

• Aider - Alternative terminal-based tool
• OpenCode - Modern TUI with similar features
• Responses API - API documentation for custom integrations