Compare
OpenAI Codex vs Bernstein
OpenAI Codex CLI is OpenAI's terminal coding agent built on the GPT-5.5 family. The `codex exec` command runs non-interactively with `--full-auto` and emits structured JSON so Bernstein parses every step.
Last verified against @openai/codex 0.117.x on 2026-05-05.
Install both
OpenAI Codex
npm i -g @openai/codex (or brew install --cask codex)Bernstein
pipx install bernsteinApache-2.0 · 42 adapters · Python scheduler · zero LLM tokens spent on coordination.
Feature matrix
| Capability | OpenAI Codex | Bernstein |
|---|---|---|
| License | OpenAI proprietary | Apache-2.0 |
| Authentication | OPENAI_API_KEY | Per-agent credential scoping (no global key) |
| Parallel-safe in worktrees | Yes | Yes (designed around git worktrees) |
| MCP support | Yes | Yes (MCP server + tool injection) |
| HMAC-chained audit log | No | Yes (RFC 2104 SHA-256 chain in .sdd/) |
| Coordinator is an LLM | — | No — deterministic Python scheduler |
| Multi-agent orchestration | One agent in a terminal | OpenAI Codex + 41 other adapters in parallel |
Source: src/bernstein/adapters/codex.py · upstream
When to choose which
Choose OpenAI Codex
If you live inside the OpenAI ecosystem (gpt-5.5 / gpt-5.5-mini / o3) and want one tool that handles plan, edit, and apply — Codex is the canonical answer. The o-series reasoning models are also accepted by the same binary, no adapter switch needed.
Choose Bernstein
When the same goal benefits from a cross-model second opinion: have Codex draft, have Claude review, have Aider apply patches that survive both. Bernstein's quality gates and cross-model verifier make this a one-line config in `bernstein.yaml`, not a hand-rolled shell script.
FAQ
Does Bernstein replace OpenAI Codex?
No. Bernstein wraps OpenAI Codex as one of 42 adapters and routes tasks to it based on a per-task pass-rate history. OpenAI Codex keeps doing what it does well; the orchestrator decides when it gets the work.
What does the audit chain give me that OpenAI Codex can't?
An HMAC-SHA256 row chain over every routing decision and every quality-gate result, written under .sdd/. Deleting or rewriting a row breaks the chain and the verifier rejects the file on next read. Useful for compliance reviews where the regulator wants to replay the run.
Can I run OpenAI Codex alongside Claude Code, Codex, and Aider in the same session?
Yes. That's the design point. Each agent runs in its own git worktree, so file edits never collide. Bernstein merges results back to the branch only after the configured quality gates pass.