Hermes Agentvs Codex

Codex is OpenAI's coding agent for writing, reviewing and shipping code. It inspects files, edits code and runs the tools on your machine in the ChatGPT desktop app, the Codex CLI or an IDE extension, and Codex Cloud runs tasks on OpenAI-managed computers. Hermes Agent is the agent around that work. It answers on more than 20 messaging platforms, runs scheduled jobs, keeps memory and works with the model you pick.

A bundled Hermes Agent skill, codex, is listed in the skills catalog as 'Delegate coding to OpenAI Codex CLI (features, PRs).' Codex itself signs in with your ChatGPT account or an API key. Hermes Agent can also sign in with your ChatGPT account through its OpenAI Codex provider, which uses Codex models and can import existing Codex CLI credentials.

If you already use the Codex CLI, hermes migrate codex brings your setup into Hermes Agent, installing the official agent-importers plugin when it is missing. Your global AGENTS.md and Codex memories become memory entries, and your MCP servers and skills come along. It shows a preview before it writes anything and never reads your credential files, and hermes import-agent --sync pulls in later changes.

Hermes Agent answers on more than 20 messaging platforms, including Telegram, WhatsApp, Slack, Signal and Microsoft Teams, from one background gateway. That gateway owns every session, so the terminal, Hermes Desktop, your chat apps and scheduled jobs share one conversation. You schedule jobs in natural language or cron syntax, and on a server or Hermes Cloud the gateway keeps running them with your computer off.

In Slack you mention @ChatGPT, and when the workspace has Cloud delegation enabled, ChatGPT can send repository work to a Codex Cloud task and return the result to the thread. In Linear you assign an issue to Codex or mention @Codex, and on GitHub you comment @codex review on a pull request. Desktop scheduled tasks that need local files need the computer on and the app running, and the Codex CLI has no interface for managing scheduled tasks.

Hermes Agent keeps a curated memory of your preferences, projects and environment, and it loads that memory into every session. Full-text search covers all its past CLI and messaging sessions. When it works out a non-trivial workflow, it saves the approach as a skill and refines that skill as it uses it. You can require approval before skill changes land, and a background curator maintains the skills it creates.

Local Codex memories are off by default. When you turn them on, Codex stores them as files under ~/.codex/memories, separate from ChatGPT memory, and OpenAI says required team guidance belongs in AGENTS.md. On macOS, Record & Replay lets eligible users demonstrate a workflow once and turn it into a reusable skill.

Hermes Agent works with any OpenAI-compatible API. Nous Portal gives it hundreds of models with one login, new frontier models included the day they launch, and local models run through Ollama, vLLM, llama.cpp or SGLang. You pick the provider and model, /model switches mid-session, and your conversation history, memory and skills carry over whichever provider you use. Hermes Agent can also use a SuperGrok sign-in or ChatGPT sign-in for Codex models.

With ChatGPT sign-in, Codex uses OpenAI models. With an API key, models follow your API organization. Local clients also have built-in providers for Amazon Bedrock, Ollama and LM Studio, and custom providers need a Responses API endpoint; current Codex releases do not support Chat Completions-only endpoints.

Hermes Agent stores conversations, memory and skills in its home directory (~/.hermes on macOS and Linux) on the machine that runs it. It collects no telemetry by default, and usage-metrics sharing is opt-in. Model requests go only to the providers you configure, and local models run on your own computer.

By default, Codex periodically sends a small amount of anonymous usage and health data back to OpenAI, which says the metrics contain no personally identifiable information. In the Codex CLI, a config flag turns analytics and metrics off.