Hermes Agent alternatives: OpenClaw, Claude Code, Codex, Pi and more
The agents that do Hermes Agent's job differently, as of November 2026: open source or not, where each runs, models, chat apps, memory, schedules, price.
The closest alternative to Hermes Agent is OpenClaw: also free and MIT-licensed, also an always-on agent on a computer you choose, reached from Telegram, WhatsApp, Discord, Slack and many more chat apps, with any model, memory in Markdown files, self-written skills and a built-in scheduler. It imports a Hermes setup with openclaw migrate hermes. If what you actually use Hermes for is code, Claude Code and Codex now schedule work too, and Claude Code can be messaged from Telegram, Discord or iMessage, but both are built around a repository rather than a chat companion. Pi is a minimal open-source harness to build your own on. And if you would rather not run anything, the labs now run an agent on their own cloud for you: OpenAI’s dots, Meta’s Muse, Google’s Task in Gemini, Anthropic’s Claude Duet and xAI’s Grok Bot, each with its maker’s model and apps.
What you would be replacing
Hermes is Nous Research’s self-improving agent: one gateway process that carries the same agent into the terminal and 28 messaging platforms, a small curated memory (two files capped at 2,200 and 1,375 characters, plus a searchable history of past conversations), skills it writes for itself, and a cron scheduler that delivers to any of those platforms. It runs any model, including local ones through llama.cpp, and by default a second model judges each risky command before it runs. The current release is v0.23.3, from November 20, 2026. What Hermes Agent is covers it in full. So an alternative is worth weighing on the same points: whether its code is open, where it runs, which models, which chat apps, how it remembers, and what it costs.
Side by side, as of November 2026
| Hermes Agent | OpenClaw | Claude Code | Codex | Pi | |
|---|---|---|---|---|---|
Licence | MIT | MIT | Closed: “All rights reserved” | Apache-2.0 | MIT |
Where it runs | Your computer or server | Your computer or server; its makers offer no hosted version | Your computer or server; cloud scheduled tasks on Anthropic’s side | Your computer (CLI, IDE, ChatGPT desktop app); web tasks on OpenAI’s side | Your computer or server, in a terminal |
Models | Any provider, or local through llama.cpp | Any provider, local models, or a Claude or Codex harness as a plugin | Anthropic’s Claude models | OpenAI’s models | 15-plus providers, hundreds of models |
Chat apps | 28 platforms in its docs | 33 entries in its channel catalog | Telegram, Discord, iMessage, as a research preview, while the session is open | None to talk to it in; Gmail, Slack or GitHub activity can start a scheduled task | None built in |
Memory | Two capped files plus searchable past conversations | Markdown files that grow, consolidated overnight | CLAUDE.md you write, plus notes it writes per repository | AGENTS.md you write; its own memories off by default | None built in; left to extensions |
Schedules | Built-in cron, delivered to any platform | Built-in scheduler, delivered to a chat or a webhook | In-session loops; Desktop tasks; cloud tasks hourly at most | Scheduled tasks in the ChatGPT desktop app, which must stay open | None built in |
Price | Free; you pay for the model | Free; “no paid tier, hosted service, or token” | From Claude Pro, $20 a month, or an API key | In ChatGPT Free, Go ($8), Plus ($20) and Pro (from $100), or an API key | Free |
The open agents cost nothing to install and bill you only through whatever model you point them at. The Claude and ChatGPT plans include model use up to their limits.
OpenClaw, the closest
OpenClaw is the same kind of program: an assistant “that runs on your own computer and meets you in the channels you already use”, through one always-on gateway. Its channel catalog holds 33 entries, among them Telegram, WhatsApp, Signal, iMessage, Discord, Slack, Microsoft Teams, Matrix and LINE. Its memory is plain Markdown files in the agent’s workspace, which grow rather than staying capped, and a background pass it calls dreaming consolidates them. It writes its own skills by default, and its scheduler wakes the agent for one-shot and recurring jobs and delivers the result to a chat. It has no paid tier; the latest release is 2026.11.6, from November 19, 2026. What OpenClaw is starts from the beginning.
The difference that matters most day to day is the default stance. Hermes starts guarded: its default approval mode has a second model approve low-risk commands, refuse dangerous ones and ask you about the rest. OpenClaw’s docs say “Sandboxing and exec approvals are off by default”: it trusts its single owner, and hardening is yours to switch on. The rest (who is behind each project, how each remembers, one agent against several) is set out in OpenClaw against Hermes.
Moving across is one command. OpenClaw’s Hermes importer previews first and then applies:
openclaw migrate hermes --dry-run
openclaw migrate apply hermes --yesCopiedIt brings the model and provider settings, MCP servers, SOUL.md and AGENTS.md, both memory files and every skill with a SKILL.md; API keys come only with your consent. Scheduled jobs, sessions, plugins and the chat-platform pairings are copied to a report folder for you to set up again rather than loaded. Run the two agents on different chat accounts if you keep both for a while.
Claude Code and Codex, if the work is code
If what you run Hermes for is mostly software work, the coding agents now schedule and take messages too, though their home is a repository and a terminal rather than your chat apps.
Claude Code is Anthropic’s, closed source, and comes with Claude Pro ($20 a month), Max or Ultra, or an API key. Its channels, a research preview, let you message a running session from Telegram, Discord or iMessage and get the reply in the same chat; Anthropic’s note is that events “only arrive while the session is open”, so on a server it runs in a persistent terminal. It schedules work three ways: /loop inside an open session (recurring tasks expire after seven days), Desktop tasks on your own machine, and cloud tasks that run on Anthropic’s side with the machine off, from a fresh clone of the repository and no more often than hourly. Its memory is the CLAUDE.md files you write plus notes it keeps per repository.
Codex is OpenAI’s, open source under Apache-2.0, and comes with every ChatGPT plan from Free upward, or runs on an API key in the CLI. Scheduled tasks are managed in the ChatGPT desktop app, and for a local project OpenAI says to “keep the machine powered on and the ChatGPT desktop app running”; the CLI has no scheduling interface of its own. Activity in Gmail, Slack or GitHub can start a task on eligible plans, and results land in the app, not in a chat. Its local memories are off by default.
Neither is a personal assistant in Hermes’s sense: their memory serves the work in a repository, and neither reaches you on WhatsApp or Signal.
Pi, if you want to build your own
Pi is the opposite of Hermes’s everything-included design: an MIT-licensed “minimal agent harness” for the terminal with 15-plus model providers and, on purpose, no permission prompts, sub-agents or plan mode. Its home page mentions no messaging apps or scheduler, and leaves long-term memory to extensions. People choose it to assemble exactly the agent they want from extensions and packages, with tmux or cron of their own around it. What Pi is covers what it leaves out, why, and where its makers advise running it, since it never asks before acting.
The labs’ hosted agents
These are closed products on a computer the lab runs, thinking with the lab’s model and reached in the lab’s apps. Nothing to install; nothing to choose either.
- OpenAI’s dots: “lives in the cloud and has its own computer and browser”; included with ChatGPT Pro (from $100 a month, not in the EEA, the UK or Switzerland) and some Business and Enterprise workspaces.
- Meta’s Muse: runs on “its own dedicated Secure Virtual Machine”, free for most use with paid plans for more; offered in the US, Canada, Japan, Singapore and, since November 17, Australia.
- Google’s Task in Gemini (formerly Spark): free in the US, Canada, Australia, India, Japan and South Korea, with Google AI Pro or Ultra elsewhere, and not available in the EEA, the UK, Switzerland or Nigeria.
- Anthropic’s Claude Duet: a virtual machine with its own browser, on Claude Opus 5.5, in a waitlisted research preview for US Pro, Max and Ultra users.
- xAI’s Grok Bot: by xAI’s FAQ, one cloud computer per account shared by all your bots (its product page says otherwise), the model picked for each task with “no model picker”, with Premium, Super or Ultra on The X Subscription.
Memory in each of them is the vendor’s to keep; with Hermes or OpenClaw it is files on a disk you control. Dots and Muse against an agent of your own and OpenClaw, Hermes and Grok Bot go row by row. If what you want is reminders and briefings from your phone more than an agent with a computer, the everyday assistant apps are the cheaper comparison.
If the problem is running it, not Hermes itself
Some searches for an alternative are really about the server: the install, the updates, keeping the dashboard off the internet. Every self-hosted agent above carries the same work, so switching agents does not remove it. If you keep running Hermes yourself, installing Hermes on a Linux server is the setup, and the same question for OpenClaw is in OpenClaw alternatives.