elonmusk / README.md
GGSheng's picture
feat: deploy Gemma 4 to hf space
3b47d98 verified
---
title: elonmusk
emoji: 🦞
colorFrom: blue
colorTo: indigo
sdk: docker
sdk_version: 29.0.4
python_version: 3.14.4
app_port: 7860
app_file: mian.py
pinned: false
---
# OpenClaw on Hugging Face Space (Docker)
> **Languages:** [English](./README.md) · [简体中文](./README_zh.md)
> **Deployment Guide:** [DEPLOY_GUIDE.md](./DEPLOY_GUIDE.md) | [中文部署指南](./DEPLOY_GUIDE_zh.md)
This setup is designed to provide the following:
- Build the OpenClaw container on top of `ubuntu:24.04`
- Serve the OpenClaw dashboard directly on port `7860` (default Space access port)
- Use third-party OpenAI-compatible `base_url + api_key` by default (injected via environment variables)
- Store OpenClaw config/workspace under `/root/.openclaw`
- Restore state automatically from a Hugging Face Dataset on startup
- Run scheduled backups of OpenClaw data to a Hugging Face Dataset via `cron` (as `root` user)
- Preinstall `python3`, `uv`, `vim`, `neovim`, `chromium` (via Chrome for Testing archive), `gh`, `hf`, `opencode`, `codex`, `claude` (Claude Code CLI), `@larksuite/cli` (with `npx skills add larksuite/cli -y -g`), and `sshx` in the image for interactive terminal use
## Repository Layout
- `Dockerfile`: Runtime image for the Space
- `scripts/openclaw-entrypoint.sh`: Main startup flow (restore, config generation, cron setup, gateway start)
- `openclaw_hf/backup.py`: Backup/restore implementation
- `scripts/openclaw-backup-cron.sh`: Cron entrypoint for backup jobs
- `scripts/openclaw-restore.sh`: Startup restore entrypoint
- `scripts/openclaw-gateway-restart`: Kill running `openclaw-gateway` processes
- `scripts/bootstrap-hf.sh`: Interactive bootstrap for Space/Dataset creation, upload, and Space variables/secrets setup (macOS/Linux)
- `scripts/bootstrap-hf.ps1`: Interactive bootstrap for Space/Dataset creation, upload, and Space variables/secrets setup (Windows PowerShell)
- `tests/test_backup.py`: Unit tests for the backup module
- `tests/test_entrypoint_config.py`: Unit tests for gateway config generation behavior
## Required Variables (Space Settings)
In your Hugging Face Space (`Settings -> Variables and secrets`), configure at least:
- Variable: `OPENCLAW_BACKUP_DATASET_REPO`: Backup target Dataset in `username/dataset-name` format
- Secret: `HF_TOKEN`: Used to write backups to the Dataset (must have write permission to that Dataset)
- Secret: `OPENCLAW_GATEWAY_TOKEN`: Gateway token (recommended; if omitted in deployment workflow, generate a random 32-character value)
- Secret: `OPENCLAW_GATEWAY_PASSWORD`: Gateway password (optional; if omitted in deployment workflow, generate a random 16-character value)
When using `./scripts/bootstrap-hf.sh` (macOS/Linux) or `./scripts/bootstrap-hf.ps1` (Windows PowerShell), these values are configured automatically on the target Space.
## Optional LLM Variables (All-Or-None)
Set all of these together only when you want OpenClaw to preconfigure a custom third-party model:
- Variable: `OPENCLAW_LLM_BASE_URL`: Third-party base URL (for example OpenAI-compatible `/v1`)
- Variable: `OPENCLAW_LLM_MODEL`: Third-party model ID
- Secret: `OPENCLAW_LLM_API_KEY`: Third-party API key
If any of the three is missing, entrypoint skips custom model generation.
In that case, you can still configure from inside the container (for example via `sshx`).
## Common Optional Variables
- `OPENCLAW_LLM_MODEL` (unset by default; used only when custom model preconfiguration is enabled)
- `OPENCLAW_LLM_PROVIDER` (default: `thirdparty`)
- `OPENCLAW_LLM_API` (default: `openai-completions`)
- `OPENCLAW_VERSION` (used by Docker install step; bootstrap prompts for it and defaults to the latest version detected from npm registry, fallback `latest`)
- `OPENCLAW_STATE_DIR` (default: `/root/.openclaw`)
- `OPENCLAW_USER` (default: `root`, runtime user for gateway and cron jobs)
- `OPENCLAW_GROUP` (default: `root`, runtime group for gateway and cron jobs)
- `OPENCLAW_CONFIG_PATH` (default: `/root/.openclaw/openclaw.json`)
- `OPENCLAW_WORKSPACE_DIR` (default: `/root/.openclaw/workspace`)
- `OPENCLAW_BACKUP_CRON` (default: `*/30 * * * *`, backup every 30 minutes)
- `OPENCLAW_BACKUP_SOURCE_DIR` (default: `/root/.openclaw`, backup/restore base directory for `openclaw-state`)
- `OPENCLAW_BACKUP_ROOT_CONFIG_DIR` (default: `/root/.config`, additional backup/restore directory for `root-config`)
- `OPENCLAW_BACKUP_ROOT_CODEX_DIR` (default: `/root/.codex`, additional backup/restore directory for `root-codex`)
- `OPENCLAW_BACKUP_ROOT_CLAUDE_DIR` (default: `/root/.claude`, additional backup/restore directory for `root-claude`)
- `OPENCLAW_BACKUP_ROOT_AGENTS_DIR` (default: `/root/.agents`, additional backup/restore directory for `root-agents`)
- `OPENCLAW_BACKUP_ROOT_SSH_DIR` (default: `/root/.ssh`, additional backup/restore directory for `root-ssh`)
- `OPENCLAW_BACKUP_ROOT_ENV_DIR` (default: `/root/.env.d`, additional backup/restore directory for `root-env`)
- `OPENCLAW_BACKUP_ROOT_NPM_DIR` (default: `/root/.npm`, additional backup/restore directory for `root-npm`)
- `OPENCLAW_BACKUP_ROOT_LARK_CLI_DIR` (default: `/root/.lark-cli`, additional backup/restore directory for `root-lark-cli`)
- `OPENCLAW_BACKUP_PATH_PREFIX` (default: `backups`)
- `OPENCLAW_BACKUP_KEEP_COUNT` (default: `24`, keep newest N timestamped backup archives; older ones are auto-deleted)
- `OPENCLAW_SSHX_AUTO_START` (default: `false`; set `true` to auto-run `sshx` in background on startup)
- `OPENCLAW_GATEWAY_AUTH_MODE` (default: `token`, optional: `password`)
- `OPENCLAW_GATEWAY_CONTROLUI_ALLOW_INSECURE_AUTH` (default: `false`)
- `OPENCLAW_GATEWAY_CONTROLUI_DANGEROUSLY_DISABLE_DEVICE_AUTH` (default: `false`; set `true` to bypass pairing, strongly discouraged on public networks)
## Quick Deployment
Run the interactive bootstrap script from repo root:
```bash
./scripts/bootstrap-hf.sh
```
```powershell
powershell -ExecutionPolicy ByPass -File .\scripts\bootstrap-hf.ps1
```
`bootstrap-hf.sh` / `bootstrap-hf.ps1` will:
- Check/install `hf` CLI:
- macOS/Linux: `curl -LsSf https://hf.co/cli/install.sh | bash`
- Windows PowerShell: `powershell -ExecutionPolicy ByPass -c "irm https://hf.co/cli/install.ps1 | iex"`
- Resolve HF auth first (before all other variables):
- if `hf auth whoami` is not logged in: prompt `HF_TOKEN` and run `hf auth login --token <HF_TOKEN>`
- if already logged in: ask whether to use current user
- choose `yes`: continue
- choose `no`: backup current token, prompt new `HF_TOKEN`, run `hf auth login --token <HF_TOKEN>`, and restore the previous token at the end
- Ask for `space_name`, `dataset_name`, `OPENCLAW_VERSION`, gateway token/password, and optional LLM settings
- Default `OPENCLAW_VERSION` to latest detected from npm registry (`openclaw`), fallback `latest` when detection fails
- Auto-generate `OPENCLAW_GATEWAY_TOKEN` (32 chars) and `OPENCLAW_GATEWAY_PASSWORD` (16 chars) if left empty
- Create private Space + Dataset and upload this repository
- Configure Space `Variables and secrets` automatically, including:
- `OPENCLAW_BACKUP_DATASET_REPO`
- `OPENCLAW_VERSION`
- `HF_TOKEN`
- `OPENCLAW_GATEWAY_TOKEN`
- `OPENCLAW_GATEWAY_PASSWORD`
- `OPENCLAW_GATEWAY_CONTROLUI_ALLOW_INSECURE_AUTH=false`
- `OPENCLAW_GATEWAY_CONTROLUI_DANGEROUSLY_DISABLE_DEVICE_AUTH=false`
- Optionally configure LLM triplet and set `OPENCLAW_SSHX_AUTO_START` from prompt choice (`true`/`false`)
- Print planned deployment settings and require a final confirmation before creating/updating Space/Dataset resources
- Print Hugging Face Space page URL, app URL, and `/healthz`
If gateway token/password were auto-generated, the script prints them at the end.
## Agent Hand-off Prompt
Copy and send to your agent:
```
Please deploy OpenClaw to Hugging Face by strictly following the deployment skill in https://github.com/tenfyzhong/openclaw-hf/blob/main/SKILL.md
```
## Hugging Face Keep-Alive
How to keep a Space available depends on hardware tier:
- Free `cpu-basic`: the Space sleeps after inactivity (currently around 48h). It cannot be configured to run forever on free hardware.
- Paid hardware: the Space runs continuously by default. In `Settings -> Hardware`, set `Sleep time` to `Never` (or use API with `sleep_time=-1`) for true 24/7 availability.
- Cost-saving mode on paid hardware: set a custom `Sleep time` (for example `3600` seconds) so it auto-sleeps and auto-wakes on the next visit.
Space URL composition:
- Space repo ID format: `<owner>/<space_name>` (example: `tenfyzhong/openclaw-hf`)
- Public runtime host format: `https://<owner>-<space_name>.hf.space`
- OpenClaw health check URL: `https://<owner>-<space_name>.hf.space/healthz`
- Inside the Space runtime, Hugging Face also provides `SPACE_HOST`, so health URL can be built as `https://${SPACE_HOST}/healthz`.
Example:
```bash
OPENCLAW_HF_SPACE_ID="tenfyzhong/openclaw-hf"
SPACE_HOST="${OPENCLAW_HF_SPACE_ID/\//-}.hf.space"
HEALTH_URL="https://${SPACE_HOST}/healthz"
echo "$HEALTH_URL"
```
Keep-alive by periodic health checks:
```bash
*/12 * * * * HF_TOKEN=hf_xxx /path/to/repo/scripts/check-space-health.sh tenfyzhong/openclaw-hf >/dev/null || true
```
Notes:
- For private Spaces, unauthenticated calls to `https://<owner>-<space_name>.hf.space/healthz` return a Hub 404 page. This is expected access control behavior.
- For private Spaces, include `Authorization: Bearer <HF_TOKEN>` (the helper script above does this automatically via `HF_TOKEN` or `HUGGINGFACE_HUB_TOKEN`).
- This ping strategy is a practical workaround for reducing idle sleep on free hardware, but it is not a guaranteed always-on method.
- If you need strict 24/7 uptime, use paid hardware and set sleep time to `Never`.
References:
- <https://huggingface.co/docs/hub/spaces-gpus#sleep-time>
- <https://huggingface.co/docs/huggingface_hub/package_reference/space_runtime>
- <https://huggingface.co/docs/hub/spaces-overview>
Programmatic options (owner token required):
```python
from huggingface_hub import HfApi
api = HfApi(token="hf_xxx")
repo_id = "your-username/your-space"
# Keep running (paid hardware)
api.set_space_sleep_time(repo_id=repo_id, sleep_time=-1)
# Or sleep after 1 hour of inactivity
api.set_space_sleep_time(repo_id=repo_id, sleep_time=3600)
# Manual control
api.pause_space(repo_id=repo_id)
api.restart_space(repo_id=repo_id)
```
For this project, if you need stable dashboard access without cold starts, use paid hardware and set sleep time to `Never`.
## Backup/Restore Flow
- Startup restore: on container startup, it fetches `latest-backup.json` from the Dataset to locate the latest backup and restore:
- `openclaw-state` -> `OPENCLAW_BACKUP_SOURCE_DIR` (default `/root/.openclaw`)
- `root-config` -> `OPENCLAW_BACKUP_ROOT_CONFIG_DIR` (default `/root/.config`, restored only when present in archive)
- `root-codex` -> `OPENCLAW_BACKUP_ROOT_CODEX_DIR` (default `/root/.codex`, restored only when present in archive)
- `root-claude` -> `OPENCLAW_BACKUP_ROOT_CLAUDE_DIR` (default `/root/.claude`, restored only when present in archive)
- `root-agents` -> `OPENCLAW_BACKUP_ROOT_AGENTS_DIR` (default `/root/.agents`, restored only when present in archive)
- `root-ssh` -> `OPENCLAW_BACKUP_ROOT_SSH_DIR` (default `/root/.ssh`, restored only when present in archive)
- `root-env` -> `OPENCLAW_BACKUP_ROOT_ENV_DIR` (default `/root/.env.d`, restored only when present in archive)
- `root-npm` -> `OPENCLAW_BACKUP_ROOT_NPM_DIR` (default `/root/.npm`, restored only when present in archive)
- `root-lark-cli` -> `OPENCLAW_BACKUP_ROOT_LARK_CLI_DIR` (default `/root/.lark-cli`, restored only when present in archive)
- Scheduled backup: cron runs based on `OPENCLAW_BACKUP_CRON`
- Shutdown backup: when the container receives a stop signal, one final backup is uploaded before exit
- Each backup upload includes:
- `backups/openclaw-backup-<timestamp>.tar.gz`
- archive root `openclaw-state/`
- archive root `root-config/` (included when `OPENCLAW_BACKUP_ROOT_CONFIG_DIR` exists)
- archive root `root-codex/` (included when `OPENCLAW_BACKUP_ROOT_CODEX_DIR` exists)
- archive root `root-claude/` (included when `OPENCLAW_BACKUP_ROOT_CLAUDE_DIR` exists)
- archive root `root-agents/` (included when `OPENCLAW_BACKUP_ROOT_AGENTS_DIR` exists)
- archive root `root-ssh/` (included when `OPENCLAW_BACKUP_ROOT_SSH_DIR` exists)
- archive root `root-env/` (included when `OPENCLAW_BACKUP_ROOT_ENV_DIR` exists)
- archive root `root-npm/` (included when `OPENCLAW_BACKUP_ROOT_NPM_DIR` exists)
- archive root `root-lark-cli/` (included when `OPENCLAW_BACKUP_ROOT_LARK_CLI_DIR` exists)
- `latest-backup.json`
- Retention: after upload, only the newest `OPENCLAW_BACKUP_KEEP_COUNT` timestamped archives are kept (default `24`); older timestamped archives are deleted automatically
## Use sshx Inside the Container
`sshx` is preinstalled in the image.
1. Auto-start `sshx` in background via environment variables:
```bash
OPENCLAW_SSHX_AUTO_START=true
```
When enabled, entrypoint starts `sshx` in background and sends `sshx` output directly to container stdout/stderr logs (no file logging).
2. Manual start inside container:
```bash
sshx
```
3. Let OpenClaw start a process itself (run in OpenClaw terminal/tool):
```bash
nohup sshx >/proc/1/fd/1 2>/proc/1/fd/2 &
```
4. After use, close `sshx` process promptly:
```bash
pgrep -fa sshx
pkill -TERM -f '(^|/)sshx($| )'
```
## Local Test
```bash
python3 -m unittest discover -s tests -p 'test_*.py'
```
Pull Requests to `main` run GitHub Actions CI automatically (`.github/workflows/pr-ci.yml`):
- Unit tests: `python3 -m unittest discover -s tests -p 'test_*.py'`
- Docker image build: `docker build` (via Buildx) with `OPENCLAW_VERSION=latest`
## License
MIT. See `LICENSE`.