Self-hosted Hermes AI agent gateway for Hugging Face Spaces. HuggingMes runs Nous Research Hermes Agent on HuggingFace Spaces, giving you a 24/7 personal AI assistant with a management dashboard, persistent HF Dataset backup, and automatic connectivity fixes for blocked outbound traffic. HuggingMes directly wires the startup providers listed below, and it can also use Hermes providers configured through hermes model or config.yaml.
Table of Contents
- ✨ Features
- 🎥 Video Tutorial
- 🚀 Quick Start
- 🔐 Access Control
- 🤖 LLM Providers
- 📱 Telegram Setup
- 🌐 Cloudflare Proxy
- 💾 Backup & Persistence
- 📦 Ephemeral Package Re-install
- 🔑 API Key Rotation
- 💓 Staying Alive
- 🔐 Security & Advanced
- 💻 Terminal Access (JupyterLab)
- 🏗️ Architecture
- 🐛 Troubleshooting
- 🌟 More Projects
✨ Features
- 🧠 Hermes Core: Runs Hermes Agent for multi-turn chat, tools, memory, and agent workflows.
- 🔐 Secure by Default: Protects the dashboard and API with a single gateway token.
- 🌐 Built-in Connectivity: Adds Cloudflare Worker proxy support for Telegram and other blocked outbound traffic.
- 📊 Dashboard: Real-time view of uptime, sync health, model, provider, and agent status at
/. - 💾 Persistent Backup: Syncs chats, config, and session data to a private HF Dataset.
- ⏰ Keep-Alive: Can provision a cron-triggered Cloudflare Worker to keep the Space awake.
- 💻 Terminal (opt-in): JupyterLab terminal at
/terminal/whenDEV_MODE=trueandGATEWAY_TOKENis set. - 🔄 Self-Healing Gateway: Gateway, dashboard, health server, and JupyterLab are all monitored and automatically restarted if they exit unexpectedly.
- 📦 Ephemeral Package Replay: Install packages from the terminal and they survive restarts — shell wrappers record
apt/pip/uv/npm/hermesinstalls and replay them on every boot. - 🚀 Startup Scripts: Run arbitrary bash at boot via
HUGGINGMES_RUNorHUGGINGMES_APT/PIP/NPM_PACKAGESvariables. - 🔑 API Key Pool Rotation: Supply comma-separated key pools (e.g.
ANTHROPIC_API_KEYS=key1,key2) and the first key is promoted automatically. - 🤖 Broad Provider Support: Supports Hermes' native providers, direct API-key providers, OAuth providers, and custom OpenAI-compatible endpoints.
🎥 Video Tutorial
Watch a quick walkthrough on YouTube: Deploying HuggingMes on HF Spaces.
🚀 Quick Start
Step 1: Duplicate this Space
Step 2: Add Your Secrets
In your Space's Settings → Variables and secrets, add these under Secrets:
LLM_API_KEY- Your provider API key for direct providers.LLM_MODEL- The model ID to use, such asopenrouter/anthropic/claude-sonnet-4,openai/gpt-4o, orgoogle/gemini-2.5-flash.GATEWAY_TOKEN- A strong token to secure the dashboard.TELEGRAM_BOT_TOKEN- Telegram bot token from BotFather.TELEGRAM_ALLOWED_USERS- Comma-separated numeric Telegram user IDs.HF_TOKEN- Hugging Face token with write access for backups and HF providers.CLOUDFLARE_WORKERS_TOKEN- Cloudflare token for outbound proxying and keep-alive automation.
Step 3: Deploy & Run
After the Space builds, open it and click Open Hermes UI to access the agent interface.
🔐 Access Control
Hermes' built-in dashboard is wrapped by HuggingMes:
- Dashboard: Opening
/app/requiresGATEWAY_TOKEN. - API: Routes under
/v1/*requireAuthorization: Bearer <GATEWAY_TOKEN>.
🤖 LLM Providers
HuggingMes supports Hermes providers in two different ways:
- Direct startup providers: Set
LLM_API_KEYandLLM_MODEL, and HuggingMes maps them during boot. - Hermes-native providers: Use
hermes modelafter the Space starts, or editconfig.yamlthrough the Hermes UI. - Custom OpenAI-compatible endpoints: Point Hermes at your own
/v1endpoint.
Direct startup providers
These are the providers that HuggingMes maps directly from LLM_MODEL and LLM_API_KEY at startup.
| Provider | Prefix | Example LLM_MODEL |
Key env |
|---|---|---|---|
| OpenRouter | openrouter/ |
openrouter/anthropic/claude-sonnet-4 |
LLM_API_KEY -> OPENROUTER_API_KEY |
| Hugging Face Inference Providers | huggingface/ or hf/ |
huggingface/Qwen/Qwen3-235B-A22B-Thinking-2507 |
LLM_API_KEY -> HF_INFERENCE_TOKEN (set HF_TOKEN separately for backup) |
| AI Gateway / Vercel AI Gateway | ai-gateway/ or vercel-ai-gateway/ |
ai-gateway/openai/gpt-4o |
LLM_API_KEY -> AI_GATEWAY_API_KEY |
| Anthropic | anthropic/ |
anthropic/claude-sonnet-4-6 |
LLM_API_KEY -> ANTHROPIC_API_KEY |
| OpenAI / OpenAI Codex | openai/ or openai-codex/ |
openai/gpt-4o |
LLM_API_KEY -> OPENAI_API_KEY |
| Google Gemini | google/ or gemini/ |
google/gemini-2.5-flash |
LLM_API_KEY -> GOOGLE_API_KEY and GEMINI_API_KEY |
| DeepSeek | deepseek/ |
deepseek/deepseek-chat |
LLM_API_KEY -> DEEPSEEK_API_KEY |
| Kimi / Moonshot | kimi-coding/ or moonshot/ |
kimi-coding/kimi-for-coding |
LLM_API_KEY -> KIMI_API_KEY |
| Kimi / Moonshot (China) | kimi-coding-cn/ |
kimi-coding-cn/kimi-k2.5 |
LLM_API_KEY -> KIMI_CN_API_KEY |
| Arcee AI | arcee/ |
arcee/trinity-large-thinking |
LLM_API_KEY -> ARCEEAI_API_KEY |
| GMI Cloud | gmi/ |
gmi/zai-org/GLM-5.1-FP8 |
LLM_API_KEY -> GMI_API_KEY |
| MiniMax | minimax/ |
minimax/MiniMax-M2.7 |
LLM_API_KEY -> MINIMAX_API_KEY |
| MiniMax (China) | minimax-cn/ |
minimax-cn/MiniMax-M2.7 |
LLM_API_KEY -> MINIMAX_CN_API_KEY |
| Alibaba Cloud | alibaba/ |
alibaba/qwen3.5-plus |
LLM_API_KEY -> DASHSCOPE_API_KEY |
| Alibaba Coding Plan | alibaba-coding-plan/ |
alibaba-coding-plan/qwen3-coder-plus |
LLM_API_KEY -> DASHSCOPE_API_KEY |
| Xiaomi MiMo | xiaomi/ |
xiaomi/mimo-v2-pro |
LLM_API_KEY -> XIAOMI_API_KEY |
| Tencent TokenHub | tencent-tokenhub/ |
tencent-tokenhub/hy3-preview |
LLM_API_KEY -> TOKENHUB_API_KEY |
| Z.ai / GLM | zai/, z-ai/, z.ai/, or glm/ |
zai/glm-5 |
LLM_API_KEY -> GLM_API_KEY |
| NVIDIA NIM | nvidia/ |
nvidia/nemotron-3-super-120b-a12b |
LLM_API_KEY -> NVIDIA_API_KEY |
| xAI / Grok | xai/ or grok/ |
xai/grok-4-1-fast-reasoning |
LLM_API_KEY -> XAI_API_KEY |
| Kilo Code | kilocode/ |
kilocode/<model-id> |
LLM_API_KEY -> KILOCODE_API_KEY |
| OpenCode Zen | opencode-zen/ |
opencode-zen/<model-id> |
LLM_API_KEY -> OPENCODE_ZEN_API_KEY |
| OpenCode Go | opencode-go/ |
opencode-go/<model-id> |
LLM_API_KEY -> OPENCODE_GO_API_KEY |
Hermes-native providers and OAuth flows
These providers are supported by Hermes and can be used in HuggingMes once the agent config is set through hermes model or config.yaml. HuggingMes does not auto-map them from LLM_MODEL at boot unless Hermes itself handles that provider.
| Provider | How to use | Notes |
|---|---|---|
| Nous Portal | hermes model |
Subscription-based OAuth provider in Hermes |
| OpenAI Codex | hermes model |
ChatGPT OAuth / Codex models |
| GitHub Copilot | hermes model |
Uses COPILOT_GITHUB_TOKEN, GH_TOKEN, or gh auth token |
| GitHub Copilot ACP | hermes model |
Spawns the Copilot CLI backend |
| Anthropic (OAuth / Claude Code) | hermes model |
Also supports ANTHROPIC_API_KEY |
| Google Gemini (OAuth) | hermes model |
Browser OAuth flow, including free-tier Gemini OAuth |
| Qwen Portal (OAuth) | hermes model |
Alibaba Qwen portal OAuth login |
| MiniMax (OAuth) | hermes model |
Portal login for MiniMax models |
| Hugging Face Inference Providers | hermes model |
Unified HF provider routing with model suffixes like :fastest and :cheapest |
| AWS Bedrock | hermes model or config.yaml |
Uses AWS credentials chain, not an API key |
| Ollama Cloud | hermes model |
Managed Ollama catalog with OLLAMA_API_KEY |
| Arcee AI | hermes model |
First-class Hermes provider |
| GMI Cloud | hermes model |
First-class Hermes provider |
| Alibaba Cloud / DashScope | hermes model |
First-class Hermes provider for Qwen models |
| Tencent TokenHub | hermes model |
First-class Hermes provider |
| Custom endpoint | hermes model or config.yaml |
Any OpenAI-compatible endpoint |
Custom and self-hosted endpoints
HuggingMes also works with any OpenAI-compatible server. Common examples include local Ollama, LM Studio, llama.cpp / llama-server, vLLM, SGLang, LocalAI, Jan, LiteLLM, ClawRouter, Together AI, Groq, Fireworks AI, Azure OpenAI, and similar services.
Use either the Hermes model wizard or a direct config.yaml entry with a base_url, model, and optional API key. For local servers that do not require auth, leave the key empty.
Recommended provider choices
- Just want it to work: OpenRouter or Hermes' Nous Portal.
- Want local models: Ollama, LM Studio, llama.cpp, vLLM, or SGLang through a custom endpoint.
- Need cloud APIs: OpenAI, Anthropic, Google Gemini, DeepSeek, xAI, Hugging Face, or any other direct provider above.
- Need routing or fallback: Use a custom endpoint such as LiteLLM or ClawRouter.
📱 Telegram Setup
To use Hermes via Telegram:
- Create a bot via @BotFather: send
/newbot, follow the prompts, and copy the bot token. - Find your Telegram user ID with @userinfobot — send it any message and it replies with your numeric user ID.
- Add
TELEGRAM_ALLOWED_USERSas a comma-separated list of those numeric user IDs to restrict access. - Add
CLOUDFLARE_WORKERS_TOKENif you need automatic outbound proxying for Telegram API traffic.
🌐 Cloudflare Proxy
Hugging Face Spaces often block outbound calls to APIs used by Telegram and some provider backends. HuggingMes can provision a Cloudflare Worker proxy automatically when you add CLOUDFLARE_WORKERS_TOKEN.
Scope and trade-offs:
- The default worker allowlist covers messaging/social APIs (Telegram, Discord, Slack, WhatsApp) and common HTTP email APIs — not LLM provider domains.
- AI API traffic is not proxied unless you opt in via
CLOUDFLARE_PROXY_DOMAINS(or*for all domains — routes API keys through Cloudflare). - Proxying adds a third-party hop; keep
CLOUDFLARE_PROXY_URLand worker secrets private. - Telegram webhooks hit
/telegramon your Space (noGATEWAY_TOKENat the proxy layer). HuggingMes auto-generatesTELEGRAM_WEBHOOK_SECRETwhen a webhook URL is configured — keep it set so Hermes can reject forged webhook calls.
💾 Backup & Persistence (Optional)
Set HF_TOKEN with write access to enable backup. HuggingMes syncs workspace data to a private HF Dataset named {your-username}/huggingmes-backup every 180 seconds by default.
Recommended Space variables for reliable persistence:
| Variable | Recommended | Why |
|---|---|---|
HF_TOKEN |
Write token (Secret) | Must override HF's read-only system token |
SYNC_INTERVAL |
120 |
Shorter loss window on crash/restart |
SYNC_START_DELAY |
5 |
First upload sooner after boot |
Verify backup before trusting restarts:
- Open the dashboard
/— Backup tile should showSUCCESS,RESTORED, orSYNCED(notFRESH,ERROR, or stuck onCONFIGURED). - Visit
https://huggingface.co/datasets/{your-username}/huggingmes-backup— expectsessions/,config.yaml, and recent commits. - Authenticated
/status(requiresGATEWAY_TOKEN) shows full backup details.
| Variable | Default | Description |
|---|---|---|
HF_TOKEN |
— | HF token with Write access (backup only — do not conflate with inference) |
HF_INFERENCE_TOKEN |
— | Optional separate token for huggingface/ models |
BACKUP_DATASET_NAME |
huggingmes-backup |
Dataset name for backup |
SYNC_INTERVAL |
180 |
Backup frequency in seconds |
SYNC_START_DELAY |
5 |
Delay before first sync after boot |
📦 Ephemeral Package Re-install (Optional)
Install packages in the terminal and they survive Space restarts — no extra config needed. Shell wrappers record every successful apt install, pip install, uv pip install, npm install -g, and hermes plugins install into workspace/startup.sh, which is backed up and replayed automatically on next boot.
For packages you want installed from day one (before the terminal is even opened), use the startup variables:
| Variable | What to put in it |
|---|---|
HUGGINGMES_RUN |
Full bash script to run on every startup (multi-line, heredocs, if blocks all work) |
HUGGINGMES_APT_PACKAGES |
Space-separated apt packages to install |
HUGGINGMES_PIP_PACKAGES |
Space-separated Python packages to install |
HUGGINGMES_NPM_PACKAGES |
Space-separated npm packages to install globally |
Example:
HUGGINGMES_RUN="""
pip install pandas matplotlib
npm install -g tsx
sudo apt-get install -y ffmpeg
"""
For scripts with complex quoting, base64-encode them:
# locally
base64 -w0 setup.sh
# HF Variable
HUGGINGMES_RUN=base64:<paste-output-here>
🔑 API Key Rotation (Optional)
Spread requests across multiple API keys to avoid rate limits. Supply a comma-separated pool — only the first key is promoted to the provider's singular env var at boot. HuggingMes does not rotate keys at runtime on rate-limit errors; Hermes would need to handle that itself.
ANTHROPIC_API_KEYS=sk-ant-key1,sk-ant-key2
OPENAI_API_KEYS=sk-oai-key1,sk-oai-key2
OPENROUTER_API_KEYS=sk-or-key1,sk-or-key2
Supported pool vars: OPENROUTER_API_KEYS, ANTHROPIC_API_KEYS, OPENAI_API_KEYS, GOOGLE_API_KEYS, GEMINI_API_KEYS, DEEPSEEK_API_KEYS, KIMI_API_KEYS, MINIMAX_API_KEYS, NVIDIA_API_KEYS, XAI_API_KEYS, KILOCODE_API_KEYS, GLM_API_KEYS, ARCEEAI_API_KEYS, DASHSCOPE_API_KEYS, GMI_API_KEYS, TOKENHUB_API_KEYS.
💓 Staying Alive
With CLOUDFLARE_WORKERS_TOKEN set, HuggingMes can create a keep-alive worker that pings the Space's /health endpoint on a schedule.
Limitations: This is a best-effort workaround, not a guarantee. Free-tier Spaces may still sleep; cold starts are slow; the cron Worker is an external dependency. /health always returns HTTP 200 when the health-server is up (gateway readiness is in the JSON body). Use /ready for stricter gateway+dashboard checks.
🔐 Security & Advanced (Optional)
| Variable | Default | Description |
|---|---|---|
GATEWAY_TOKEN |
required on HF Spaces | Token for dashboard, /v1/* API, and terminal gate |
JUPYTER_TOKEN |
(uses GATEWAY_TOKEN) |
Separate terminal token (optional hardening — compromise isolation) |
HF_TOKEN |
— | HF token with write access for dataset backup |
HF_INFERENCE_TOKEN |
— | HF inference token when using huggingface/ models |
SPACE_PRIVACY |
auto-detect | Set public or private to avoid privacy-detection race |
CLOUDFLARE_WORKERS_TOKEN |
— | Cloudflare API token for proxying and keep-awake |
SYNC_INTERVAL |
180 |
Backup frequency in seconds |
SYNC_START_DELAY |
5 |
Seconds before first sync after boot |
CLOUDFLARE_KEEPALIVE_ENABLED |
true |
Set false to disable keep-awake worker |
TELEGRAM_MODE |
webhook |
webhook or polling (webhook auto-configured from SPACE_HOST) |
TELEGRAM_WEBHOOK_SECRET |
auto-generated | Shared secret for Telegram webhook validation |
DEV_MODE |
false |
Set true to enable JupyterLab terminal at /terminal/ |
WEBHOOK_URL |
— | Endpoint for POST JSON restart notifications |
GATEWAY_RESTART_DELAY |
5 |
Seconds between gateway restart attempts |
GATEWAY_MAX_RESTARTS |
0 (unlimited) |
Maximum gateway restart count before container exits |
💻 Terminal Access (JupyterLab)
HuggingMes includes an optional JupyterLab terminal at /terminal/ for direct shell access to the container — useful for running hermes commands, inspecting files, installing packages, and debugging.
Setup
The terminal is off by default (DEV_MODE=false). Set DEV_MODE=true and GATEWAY_TOKEN, then click Open Terminal → on the dashboard.
Use JUPYTER_TOKEN for a separate terminal password so a compromised API token does not automatically unlock the shell (trade-off: two secrets to manage).
Security: JupyterLab grants full shell access to the container (passwordless
sudo). Keep the Space private and disable the terminal unless you need it.
What you can do
- Run
hermesCLI commands directly - Browse and edit files in the workspace
- Install Python packages with
piporuv pip - Check logs, inspect config, debug issues
💻 Local Development
docker compose up --build
# Dashboard: http://localhost:7861
# Hermes App: http://localhost:7861/app/
🏗️ Architecture
- Dashboard (
/): Real-time management and monitoring. - Hermes App (
/app/): Secure proxied access to the Hermes UI. - API (
/v1/*): Proxied OpenAI-compatible agent API. - Terminal (
/terminal/): JupyterLab terminal (DEV_MODE=truerequired). - Health Check (
/health): Liveness probe for HF and keep-alive (always 200 when health-server is up). - Readiness (
/ready): Returns 503 until gateway, dashboard, and auth are ready. - Sync Engine: Python background task for HF Dataset persistence.
🐛 Troubleshooting
- Telegram bot not responding: Ensure
CLOUDFLARE_WORKERS_TOKENis set and check logs for the proxy setup step. - Authentication failed: Clear browser cookies or use an incognito window if
GATEWAY_TOKENchanged. - Data not persisting: Ensure
HF_TOKENhas write access, setSYNC_INTERVAL=120, wait for Backup tileSUCCESS/SYNCED, then restart. - Restore failed banner: Check dataset URL on dashboard; verify
HF_USERNAMEif namespace detection fails. - GATEWAY_TOKEN required: HF Spaces fail closed without
GATEWAY_TOKEN— set it in Space secrets. - Provider not showing up: If it is a Hermes-native provider, run
hermes modeland complete the provider-specific setup there. If it is a custom endpoint, verify thebase_urlexposes/v1/modelsor/v1/chat/completions. - Space keeps sleeping: Add
CLOUDFLARE_WORKERS_TOKENto enable automatic keep-awake monitoring.
🌟 More Projects
Similar projects by @somratpro — all free, one-click deploy on HF Spaces:
| Project | What it runs | HF Space | GitHub |
|---|---|---|---|
| HuggingFlow | DeerFlow — deep research agent | Space | Repo |
| Hugging8n | n8n — workflow & automation platform | Space | Repo |
| HuggingClaw | OpenClaw — Claude Code in the browser | Space | Repo |
| HuggingClip | Paperclip — AI agent orchestration platform | Space | Repo |
| HuggingPost | Postiz — social-media scheduler | Space | Repo |
Made with ❤️ by @somratpro