--- title: "Hermes Agent vs Codex CLI: Which Coding Agent to Use (2026)" description: "Hermes Agent (Nous Research) and Codex CLI (OpenAI) compared — model flexibility, self-improving skills, pricing, and which one fits your workflow." date: 2026-05-04 updatedDate: 2026-09-08 location: San Francisco, CA – May 4th, 2026 image: /images/hermes-vs-codex-hero.jpg keywords: "hermes codex, hermes agent codex, claude hermes, hermes agent pricing, hermes vs codex, hermes agent vs codex cli, nous research hermes, openai codex cli" faq: - question: "What's the difference between Hermes Agent and Codex CLI?" answer: "Hermes Agent (Nous Research) is a self-improving CLI agent with persistent memory, automated skill creation, and 300+ supported models across providers. Codex CLI (OpenAI) is OpenAI's first-party local coding agent — tightly integrated with GPT-5 and OpenAI's tooling, but locked to OpenAI models. Hermes also ships messaging gateways (Telegram, Slack, Discord); Codex CLI is terminal-only." - question: "How much does Hermes Agent cost?" answer: "Hermes itself is open source and free. You pay for the underlying model. With cheap providers like MiniMax M2.5 ($0.30/M) or DeepSeek V3.2 (~$0.27/M) routed through Haimaker, daily costs land around $1-5. With Claude Sonnet or GPT-5.4, expect $20-100/day on heavy use." - question: "Can I use Claude with Hermes Agent?" answer: "Yes. Hermes is OpenAI-compatible and supports any provider that exposes a /v1/chat/completions endpoint. Run 'hermes model', pick Custom endpoint, and point at Anthropic, Haimaker, OpenRouter, or any other provider. Claude Sonnet 4.6 and Opus 4.5 are popular Hermes choices." --- People searching "hermes codex" or "hermes agent codex" land in a slightly confusing spot. There are two products named in that query — **Hermes Agent** (Nous Research's open-source CLI agent) and **Codex** (OpenAI's CLI coding agent, formerly known as the GPT-5 Codex tool). They occupy similar terminal real estate but solve different problems. This is a developer-honest comparison. Both have strengths. The right pick depends on what you're doing. ## The 30-second version | | Hermes Agent | Codex CLI | |---|---|---| | **Made by** | Nous Research (open source) | OpenAI | | **Models** | 300+ across any OpenAI-compatible provider | OpenAI only (GPT-5, GPT-5 Codex variants) | | **Standout feature** | Self-evolving skills, persistent memory | First-party OpenAI integration, low latency | | **Multi-platform** | CLI + Telegram, Slack, Discord, WhatsApp | CLI only | | **Price (tool)** | Free (open source) | Free (open source, paid model) | | **Price (model)** | Whatever you wire up — $0/day local to $100+/day frontier | OpenAI rates only — typically $5-50/day | | **Best for** | Long-running personal agents, model flexibility | OpenAI-native shops, low-latency interactive use | Pick **Hermes** if you want model flexibility, persistent memory across sessions, or messaging platform integration. Pick **Codex CLI** if you're OpenAI-native and value the first-party polish over portability. ## What Hermes Agent actually does Hermes is a CLI agent that **learns the longer you run it**. Two design choices make this concrete: **1. Persistent memory.** Hermes maintains state across sessions in a structured store. Restart the terminal and it remembers your project layout, your conventions, what you were working on. Most other CLI agents start from zero each time. **2. Self-evolving skills (GEPA loop).** Every 15 tool calls Hermes pauses, evaluates what it just did, and writes a "Skill Document" capturing what worked. The next time it sees a similar task, it pulls the skill instead of re-deriving the approach. The Nous team published a peer-reviewed result showing a ~40% speedup on repeat tasks after sufficient training. Plus the practical features: - **300+ supported models** through OpenAI-compatible endpoints. Anthropic, OpenAI, Google, xAI, DeepSeek, Haimaker, OpenRouter, local Ollama, vLLM. Mid-session swaps with `hermes model`. - **Messaging gateways.** A single Hermes instance can be reached from Telegram, Slack, Discord, WhatsApp, and the terminal. The same agent state and memory persist across all of them. - **Sandboxed code execution.** Tool calls run inside a Unix-socket RPC sandbox, not directly on your shell. Less risk of an `rm -rf` going wrong. - **47+ built-in tools** — file ops, shell, web search, scheduled tasks, cross-platform messaging. The trade-off: Hermes is fast-moving. Four major releases in three weeks during the March/April 2026 cycle means the surface keeps shifting. If you want a stable target, check that the tool is at least one version behind the bleeding edge before pinning. For Hermes-specific model picks, see [Best Models for Hermes Agent](/blog/best-models-for-hermes-agent/). ## What Codex CLI actually does Codex is OpenAI's terminal coding agent. It's built around their model lineup — GPT-5, GPT-5 Codex, the o-series — and the integration is tight in a way third-party tools can't match. **The pitch:** - **First-party polish.** Latency is lower than going through OpenRouter or a custom endpoint. Tool schemas are tuned for OpenAI's models specifically. Streaming behavior is smooth. - **Local-first execution.** Code reads and edits happen on your machine; Codex calls home for inference but doesn't ship the whole codebase to OpenAI. - **Strong defaults.** No real config required to get started. Drop in your OpenAI API key and it works. **The trade-off:** OpenAI only. You can't switch to Claude when GPT-5 is being stubborn. You can't switch to MiniMax when you want cheap volume. You can't run a local model for sensitive code. Some teams accept this trade for the integration polish; others don't. If your stack is already OpenAI-native and you're paying OpenAI rates anyway, Codex CLI is the lowest-friction option. If you want to mix models (most users should — see [multi-model setup](/blog/openclaw-multi-model-routing)), Hermes or OpenClaw or OpenCode is the better fit. ## Hermes Agent pricing in practice Hermes itself is free (open source on GitHub). The cost is whatever model you point it at. A few real cost points for a 24/7 always-on Hermes agent processing ~10M tokens/day: | Provider | Daily cost | Notes | |---|---|---| | **Local Qwen3.6 27B** via Ollama | $0 | Hardware investment only | | **MiniMax M2.5** via Haimaker | ~$8 | Cheap default for most agent work | | **DeepSeek V3.2** | ~$8 | Comparable to MiniMax, slightly different strengths | | **Claude Sonnet 4.6** direct | ~$110 | Frontier reasoning, expensive for high-volume | | **Haimaker auto-router** | ~$15 | Mix of MiniMax + GPT-OSS + Claude per-request | The auto-router number is what most production Hermes setups land on. ([Setup walkthrough](/blog/openclaw-auto-router-setup) — works the same with Hermes since both use OpenAI-compatible endpoints.) For a developer using Hermes interactively (not 24/7), realistic costs are much lower — typically $1-5/day on cheap models, $5-20/day on Claude or GPT-5. Codex CLI's pricing is just OpenAI's pricing. A typical interactive day on GPT-5 lands around $5-30. There's no cheap-model option to fall back to. ## Setting up Hermes with Claude (or any model) Run: ```bash hermes model ``` Pick **Custom endpoint** from the menu. Enter: - **Base URL:** `https://api.haimaker.ai/v1` (or `https://api.anthropic.com/v1`, etc.) - **Model:** `anthropic/claude-sonnet-4-6` (or whatever fully-qualified name your provider uses) Hermes stores the choice and uses it for every subsequent run. Switch any time by running `hermes model` again. There's no JSON config to edit, which is one of Hermes's nicer features compared to OpenClaw. Setting up Hermes through Haimaker gives you all 300+ models behind one API key, which is useful if you regularly switch between Claude, GPT, Gemini, and open-source models depending on the task. You can also skip the manual endpoint entry entirely: `npx -y @haimaker/connect --hermes` writes Hermes's config to talk to `api.haimaker.ai` directly, and `--codex` does the same for Codex if you land on the OpenAI-native side of this comparison. ## Which one to pick **Use Hermes if:** - You want to mix providers (cheap default + frontier fallback) - Persistent memory and self-evolving skills matter for your workflow - You want the agent reachable from Slack or Telegram, not just the terminal - You care about not being locked to one provider **Use Codex CLI if:** - You're already OpenAI-only and paying their rates - Latency and integration polish matter more than portability - You don't need cross-platform messaging - You prefer a stable, slower-moving tool over a fast-iterating one For most developers reading this in 2026, the model-flexibility argument tips toward Hermes. The cost difference of running a cheap default plus a frontier fallback is too large to leave on the table — and OpenAI doesn't have a cheap option in the same league as MiniMax M2.5 or DeepSeek V3.2. For OpenAI-native shops with existing Codex tooling, Codex CLI's first-party polish is real. Don't switch just because of model variety if you're not actually going to use the variety. RUN HERMES THROUGH HAIMAKER --- *Related: [Best Models for Hermes Agent](/blog/best-models-for-hermes-agent/), [Cheapest API for AI coding agents](/blog/cheapest-api-coding-agents), [OpenClaw multi-model setup](/blog/openclaw-multi-model-routing).*