The OpenAI Codex CLI is a powerful coding agent — but it's locked to OpenAI's Responses API. If you want to use DeepSeek, Gemini, Anthropic, Ollama, or any other provider, you're out of the box. Codex Launcher breaks that lock-in wide open.
It's an open-source tool that lets you run the real OpenAI Codex CLI and Desktop with 20+ AI providers, adds anti-loop resilience, intelligent model profiles, a GTK desktop GUI, and requires zero pip dependencies — pure Python stdlib.
OpenAI's Codex CLI is hardcoded to the Responses API schema. That means:
Codex Launcher solves this by running a bidirectional translation proxy that intercepts the Responses API schema and translates it to Chat Completions (and back). The Codex CLI thinks it's talking to OpenAI — it's actually talking to whatever provider you choose.
Run the exact codex command from your terminal. Bypasses lock-in, handles automated self-healing, and injects proxy environment configs automatically.
A graphical dashboard with real-time metrics. Dynamically adjust system variables, normalize target providers, and keep the CLI and GUI in sync. Also has a tkinter fallback for Windows.
Intercepts OpenAI Responses API schema and translates bidirectionally to Chat Completions API. Resolves available models via /v1/models and safely stores credentials before fallbacks.
5-heuristic intent router that scans malformed commands. Three-layer self-healing: model-aware tool-call budgets, null-tool spam detection, and force-finalize to skip hallucinated responses.
Monitors every transaction — measuring raw latency, token usage, and parsing integrity. Smooth token-by-token decoding and formatting.
Runs on Bash (Linux) and Command Prompt/PowerShell (Windows). Pure standard Python — zero pip dependencies.
Codex Launcher supports every tier of AI provider, from free local models to enterprise APIs:
Two breakthrough subsystems give the AI deep understanding of your project:
CodebaseIntelligence chunks your code, indexes it in a vector database, and injects semantic context into every prompt. Sub-8ms query latency.
# semantic chunk extraction
from codex.vector import CodebaseIndex
idx = CodebaseIndex("./src")
chunks = idx.query("auth handler")
# → 847 chunks indexed · latency: 6.2ms
PrecisionSynthesis embeds AST definitions directly. The generator reads semantic context for precise, structurally-aware code synthesis.
# ast-embedded code synthesis
from codex.synth import PrecisionGenerator
gen = PrecisionGenerator(ast_defs)
result = gen.synthesize("add retry logic")
# → AST-aware synthesis complete
Same tasks, same models, different tooling. Real-world metrics:
| Feature | Codex Launcher | Claude Code | OpenCode |
|---|---|---|---|
| AI Providers | 20+ (OpenAI, Anthropic, Google, DeepSeek, Ollama, OpenRouter…) | Anthropic only | OpenCode Zen + Go only |
| Protocol Support | Responses API, Chat Completions, Anthropic Messages, Command Code, Freebuff | Anthropic Messages only | Chat Completions only |
| Tool-Call Parsing | 7-format cascading parser | Native structured output | Basic, no multi-format |
| Anti-Loop Protection | Model-aware budgets, null-tool detection, force-finalize | Basic repetition detection | None |
| Self-Healing | 3-tier AI monitor | None | None |
| Token Compaction | Proactive at 80% limit, per-model learning | Basic truncation | Basic truncation |
| Multi-Account Rotation | API keys, OAuth, Freebuff accounts | Single account | Single account |
| Free Models | Freebuff (DeepSeek V4, Kimi K2.6), Google free tier, Ollama | No free tier | Limited |
| Desktop GUI | GTK (Linux) + tkinter (Windows) | CLI only | TUI (terminal UI) |
| Dependencies | Zero pip — pure Python stdlib | Node.js + npm | Go runtime |
| Offline Support | Ollama, LM Studio, vLLM, LocalAI | Requires API | Limited |
| Cost per 1K Tasks | $0 (free) – $12 (GPT-4o) | ~$30 (Claude Opus 4.7) | ~$15 – $25 |
The project page documents the full development saga — nine phases that each solved a critical production blocker:
Install and run in seconds — zero pip dependencies, pure Python:
# Clone and run git clone https://github.com/rommark/codex-launcher cd codex-launcher python codex_launcher.py --provider deepseek --model deepseek-r1 # Or with GTK desktop python codex_launcher.py --gtk --monitor # Or use any provider codex --provider gemini --model gemini-2.5-pro codex --provider anthropic --model claude-sonnet-4 codex --provider ollama --model llama3.3