> ## Documentation Index > Fetch the complete documentation index at: https://docs.clawup.org/llms.txt > Use this file to discover all available pages before exploring further. # Runtimes > Supported agent runtimes (OpenClaw, Hermes Agent) and their differences ClawUp runs your Claw inside a managed container. The engine that powers that container is called a **runtime**. Two runtimes are supported today: **OpenClaw** and **Hermes Agent**. Both are offered across Free, Basic, Pro, and Enterprise plans — the runtime choice is independent of the subscription plan. Pick a runtime when you create a Claw or a Team. You can't change it afterwards; to switch, create a new Claw and restore from the old one (within the same runtime family — see [Restore & Migration](/src/restore-and-migration)). ## Your harness, your agent, your memory A runtime is the **harness** around your agent. In closed agent platforms, the harness is opaque — the vendor controls it, hides how memory is managed, and locks you into one model and one state format. ClawUp is deliberately the opposite: * **Your harness.** Pick OpenClaw or Hermes Agent at create time. Both are open implementations; both expose the same management API (provision, stop, start, restore, chat, files, channels). You're never stuck in a proprietary scaffolding. * **Your agent.** Pick any frontier model — Anthropic, OpenAI, OpenRouter, Google, DeepSeek, Mistral, xAI — via BYOK, or go Managed and pay from your balance. The model picker reads a live registry that refreshes with each provider's catalog, so you always see the latest. * **Your memory.** Every Claw's workspace — `SOUL.md`, `IDENTITY.md`, `TEAM.md`, agent-authored files, memory database — lives in a container you control. State tar's to your OSS bucket on stop and restores atomically on start; the [Files API](/src/features) exposes the full workspace for read/write over HTTP; the whole platform is self-hostable via `docker-compose` or Kubernetes. Your agent's history is portable across models, hosts, and regions. > In a closed harness, your agent is replicable by anyone with the same model keys. In an open harness with portable memory, your agent is **yours** — with its accumulated context, preferences, and earned behaviour. The sections below walk through each runtime's specific strengths and constraints. ## At a glance | | **OpenClaw** | **Hermes Agent** | | -------------------------------- | ----------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | | **Engine** | Node.js | Python | | **Upstream** | [ClawUp OpenClaw](https://github.com/BotMesh/openclaw) | [Nous Research Hermes Agent](https://github.com/NousResearch/hermes-agent) | | **Managed default image** | `alpine/openclaw:2026.6.9` | `nousresearch/hermes-agent:v2026.7.1` | | **Best for** | ClawHub skills, broad LLM provider coverage, mature Teams & Nebula | Built-in code execution, delegation, Python tool ecosystem | | **Skills** | ClawHub catalog (one-click install per identity) | `hermes skills install` catalog | | **Tools** | MCP + ClawUp Hooks | MCP + Hermes Hooks | | **Config format** | `openclaw.json` | `config.yaml` + `.env` | | **Config reload** | Live (SIGUSR1 — no container restart) | Container restart in place | | **Channels** | Telegram, Feishu, Discord, Slack, WhatsApp, WeCom, Mattermost | Telegram, Feishu, Discord, Slack, WhatsApp, WeCom, Mattermost | | **Supported providers** | OpenAI, Anthropic, OpenRouter, Google, DeepSeek, Moonshot, Qwen, custom | Same list, plus any OpenAI-compatible endpoint via `provider: custom` + `base_url` | | **Identity / pre-seeded skills** | Available (picked at create time) | Not available — Hermes uses its own skill catalog | ## OpenClaw The default runtime and the most mature path. OpenClaw is ClawUp's Node.js agent engine; it powers every feature that originated in ClawUp (Teams, Nebula, ClawHub, runtime defaults, SOUL.md / IDENTITY.md layering). Strongest fit when you want: * **ClawHub skill packages** — pick an identity at create time and the matching skills install automatically. No Python runtime assumed. * **Live config reloads** — channel toggles, tool bindings, and model switches apply without a container restart (SIGUSR1-based in-process reload). * **Broadest docs + examples** — every tutorial on this site assumes OpenClaw unless noted. Data layout inside the runtime: ``` $HOME/.openclaw/ ├── openclaw.json # runtime config (channels, tools, provider) ├── workspace/ # user-visible files (Files API) │ ├── SOUL.md # agent persona │ ├── IDENTITY.md # identity-level prompt │ └── … # user uploads, agent-created files ├── auth-profiles.json # per-provider API keys └── sessions/, logs/ # runtime state ``` ## Hermes Agent Hermes is an upstream open-source Python agent from Nous Research. ClawUp wraps it with the same provisioning, storage, and billing as OpenClaw so that it behaves the same way from a user-visible perspective (Create Claw, Add Model, Install Tools, chat, pair channels). Strongest fit when you want: * **Built-in code execution** — Hermes ships a terminal tool, a Python sandbox, and delegation primitives directly in the agent loop. * **Python-native tool ecosystem** — skill packages are installed via `hermes skills install ` and can pull any Python dependency. * **OpenAI-compatible proxies** — `provider: custom` + `base_url` lets you point at any OpenAI-API-shaped endpoint. Data layout inside the runtime: ``` $HERMES_DATA_DIR/ (defaults to $HOME/.openclaw for ClawUp compat) ├── config.yaml # model, provider, base_url, api_key (0600) ├── .env # TELEGRAM_BOT_TOKEN, provider keys (0600) ├── workspace/ # user-visible files (Files API) ├── skills/ # hermes skill packages ├── cron/ # scheduled jobs ├── sessions/, logs/ # runtime state └── gateway_state.json # platform connection state ``` ### Differences from OpenClaw | Area | OpenClaw | Hermes Agent | | ------------------------- | ------------------------------- | -------------------------------------------------------------------- | | Identity / ClawHub skills | Selectable at create time | Not applicable — identity picker is disabled | | Config reload | SIGUSR1 in place | Container restart in place (`kill -TERM 1` inside the pod) | | Secrets in state dir | `auth-profiles.json` (JSON) | `.env` (dotenv), plus `config.yaml` for `custom` providers | | Channel token flow | Written into `openclaw.json` | Exported as env vars (`TELEGRAM_BOT_TOKEN`, etc.) at gateway startup | | CLI inside the pod | `openclaw` on `$PATH` | `/opt/hermes/.venv/bin/hermes` (venv-sourced) | | File exposure | `$STATE_DIR/workspace` (scoped) | `$HERMES_DATA_DIR/workspace` (scoped) | ### Known gaps * **Hermes on Aliyun ACK** runs behind a readiness probe tuned for Hermes's 90–120 s cold start. First-provision takes visibly longer than OpenClaw. * **Hermes self-restart on pairing** — the first unauthorized inbound message on Telegram triggers a Hermes-internal restart cycle while the allowlist settles. Subsequent messages are stable. * **Hermes teams** still reconcile through the same loop as OpenClaw teams; Nebula and `remote_send` interop is supported but has less mileage in production. ## Picking a runtime Choose OpenClaw if you want the shortest path to a ClawHub skill pack, the most-polished Teams flow, or live config reloads without container restarts. Choose Hermes Agent if you want built-in code execution, Python-based skills, or are integrating an OpenAI-compatible LLM proxy. If you're unsure, start with OpenClaw — you can always create a second Claw on Hermes later and copy workspace files across with the Files API.