Codex Configuration
Complete reference for configuring OpenAI Codex: config.toml settings, AGENTS.md project instructions, MCP servers, safety modes, and profiles.
Codex Configuration
All Codex surfaces — CLI, IDE extension, and desktop app — share a single configuration file at ~/.codex/config.toml. This means you configure once and your settings apply everywhere.
Configuration File Locations#
| Location | Scope | Notes |
|----------|-------|-------|
| ~/.codex/config.toml | Global (user-level) | Primary configuration |
| .codex/config.toml | Project-scoped | Overrides global settings (trusted projects only) |
| requirements.toml | Organization-enforced | Admin restrictions on managed machines |
Precedence Order (Highest First)#
- CLI flags (
-c key=value) - Project config (
.codex/config.toml, closest to CWD wins) - Global config (
~/.codex/config.toml)
Essential Settings#
# Model selection model = "gpt-5.4" # Safety settings approval_policy = "on-request" sandbox_mode = "workspace-write" # Web search web_search = "cached" # "cached" (default) or "live" # Review model (separate from main model) review_model = "gpt-5.3-codex"
Key Configuration Areas#
| Area | Description | Guide | |------|-------------|-------| | AGENTS.md | Project instructions and conventions | How to teach Codex your codebase | | MCP Servers | External tool integrations | Connect Figma, databases, docs | | Safety Modes | Approval and sandbox policies | Control what Codex can do |
Profiles#
Define named profiles for different workflows:
[profiles.deep-review] model = "gpt-5.3-codex" model_reasoning_effort = "high" approval_policy = "on-request" [profiles.quick-fix] model = "gpt-5.3-codex-spark" approval_policy = "never" sandbox_mode = "workspace-write"
Switch profiles from the command line:
codex --profile deep-review "Review the auth module" codex --profile quick-fix "Fix the typo in the README"
Local and Open-Source Models#
Codex supports local models via Ollama and any provider that supports the Chat Completions or Responses APIs:
# Use a local model via Ollama codex -c model_provider="oss" -c model="codellama:34b"
# config.toml for custom provider model_provider = "custom" model = "your-model-name" base_url = "http://localhost:11434/v1"
Local models may have reduced capabilities compared to GPT-5.x models. For best results with open-source models, use models with strong instruction-following abilities.
Project-scoped config files (.codex/config.toml) are only loaded when you trust the project. This prevents untrusted repositories from overriding your safety settings.
Environment Variables#
| Variable | Purpose |
|----------|---------|
| CODEX_API_KEY | API key for authentication |
| CODEX_HOME | Override the config directory (default: ~/.codex) |
| OPENAI_API_KEY | Alternative API key variable |
Sample Configuration#
For a complete annotated sample, see the official sample config.
Next Steps#
- AGENTS.md Guide — Teach Codex your project conventions
- MCP Servers — Connect external tools
- Safety Modes — Control approval and sandbox settings