What did Hlido score Codex CLI?

Codex CLI scored 85/100 (STEADY) in Hlido's independent, hands-on review.

Is Codex CLI reliable?

Codex CLI is rated STEADY (85/100) by Hlido. Codex CLI 0.149.1 — OpenAI’s terminal coding agent. CLI surface live-tested from a cold npm install: strong command breadth, sandbox subcommand, trusted-directory guard, excellent doctor — but non-interactive exec hangs silently without credentials instead of failing fast (verifi

Does any vendor pay Hlido for placement?

No. Hlido takes no money from the agents it rates — scoring weights stay private and the evidence behind every verdict is public.

Codex CLI

Name: Codex CLI review
Item: Codex CLI
Rating: 85
Author: Hlido Editor

Coding · tested 2026-08-24 · re-test due 2026-11-22 · by the Hlido desk, not the vendor

In short: A fast, capable vendor CLI with real safety ideas — held back by a failure mode its own doctor would diagnose.

4 PASS · 1 FAIL of 5 public-surface claims

Visit Codex CLI → scorecard.json

Quick answer

Codex CLI scores 85/100 (STEADY) on Hlido’s independent, hands-on test (reviewed 2026-08-24). STEADY (85) because the public surface is broad and mostly excellent — sandbox subcommand, trusted-directory guard, self-hosting as an MCP server, an actionable doctor — but the one behavior that matters most for unatten

Hlido Editor · 2026-08-24

We installed Codex CLI cold from npm (codex-cli 0.149.1) and probed its public surface. Much of it is genuinely strong: a rich command set (non-interactive exec, code review, session resume/fork), MCP management plus the ability to run Codex itself as an MCP server, a built-in sandbox subcommand, and a trusted-directory guard that refuses to run outside a trusted git repository unless explicitly overridden — a safe default we verified. Its doctor is excellent: run in our restricted sandbox it named every real problem (missing credentials, blocked egress) with a concrete fix each. But the flagship check failed: `codex exec` without credentials printed its session banner and then hung, silently, until we killed it at 40 seconds. No auth error, no timeout, no hint — while its own doctor knows exactly what is wrong. A CLI that other agents will drive non-interactively must fail fast; this one does not. What this review does NOT cover: the agentic loop requires paid credentials and was not exercised; medium confidence. Disclosure: Hlido’s review pipeline runs on Claude models from Anthropic, a direct competitor of OpenAI. The mechanical checks are reproducible by anyone; weigh the editorial layer with that knowledge.

Why STEADY

STEADY (85) because the public surface is broad and mostly excellent — sandbox subcommand, trusted-directory guard, self-hosting as an MCP server, an actionable doctor — but the one behavior that matters most for unattended and agent-driven use, failing fast on missing credentials, failed our test outright (40s silent hang, killed by timeout). That is a production-path defect on a CLI explicitly designed for non-interactive execution, and it caps the craft dimension below the VITAL band until fixed.

Public-surface checklist

PASS Version-output (required) — codex-cli 0.149.1, exit 0
PASS Help-core-usage (required) — exec/review/resume/fork/sandbox/cloud documented
PASS Mcp-management (required) — mcp list/get/add/remove + mcp-server (serve as MCP)
PASS Doctor-health (required) — exit 1 with correct, actionable findings about our sandbox
FAIL Unauthenticated-behavior (required) — silent hang, killed at 40s (exit 124) — no auth error

What it does well

Trusted-directory guard: refuses to run outside a trusted git repo without an explicit override flag (tested)
Ships a sandbox subcommand for running commands in a Codex-provided sandbox (tested help surface)
Can run itself as an MCP server (codex mcp-server) in addition to managing external ones (tested help surface)
doctor produces precise, actionable diagnostics — it correctly named every defect of our restricted sandbox (tested)
Non-interactive exec plus code-review mode designed for scripting (tested help surface)

What it fails at

Hangs instead of failing on missing credentials: `codex exec` printed a banner then hung 40+ seconds with no error until killed (tested) — its own doctor diagnoses the same condition instantly
Primary distribution page sits behind GitHub bot protection; our browser runner was CAPTCHA-blocked (prior run evidence)

Red flags

Verified: silent 40s+ hang on unauthenticated non-interactive exec — the exact path other agents would drive
Reviewer-vendor note, disclosed: Hlido’s pipeline runs on Claude models (Anthropic), a competitor of OpenAI. Mechanical checks are reproducible; weigh the editorial layer accordingly.

Best for

Developers in the OpenAI stack who want a scriptable terminal agent with session management
Teams that value an explicit sandbox boundary for agent-executed commands
Agent builders who want a coding agent addressable AS an MCP server

Not recommended for

Unattended pipelines that need deterministic fail-fast behavior on auth/config errors (verified defect)
Anyone needing a credential-free trial of the actual editing loop

Compared to

Claude Code
Claude Code failed fast without credentials (clean error, 1.3s); Codex hung 40+ seconds in the identical test. Codex counters with a built-in sandbox subcommand and can itself serve as an MCP server. Both are vendor-locked; both hide the agentic loop behind paid auth.
Aider
Aider is open-source and model-agnostic with a live-tested edit loop (92). Codex is OpenAI-native with a broader session/cloud surface but an untested loop and a verified fail-slow defect.

Agent relevance

API CLI MCP SDK Behavioral-testable

Agentic-Commerce Readiness 65/100 · INTEGRABLE

Independent readiness for agent delegation & transaction. How it’s scored · check live

Install via npm (@openai/codex); drive non-interactively with `codex exec`; run as an MCP server with `codex mcp-server`; requires OpenAI credentials. Guard automation with explicit timeouts — verified: missing credentials hang rather than error.

Evidence

— source

scorecard.json · registry · methodology

More: compare agents · best of · developer tools · incident registry

Verdict by Hlido Editor, our automated editorial system · Method: cli-tier-1+editorial-narrative-v2+claude-native · Methodology version 2026.05 · Next review due 2026-11-22

How this page was produced. The scores, claim verdicts and evidence come from automated hands-on testing of the product’s public surface. The written analysis is drafted by an AI system, and pages publish without a person reviewing each one. Hlido publishes this record and answers for it — tell us if anything here is wrong and we will correct it.

Embed this trust badge

Live, always-current independent score — free to embed on your site or README. No vendor pays for placement.

Markdown

[![Hlido trust score](https://hlido.eu/badge/openai-codex.svg)](https://hlido.eu/check/?agent=openai-codex)

HTML

<a href="https://hlido.eu/check/?agent=openai-codex"><img src="https://hlido.eu/badge/openai-codex.svg" alt="Hlido trust score"></a>