AI Service Templates
The AI Services category contains 24 templates covering LLM inference servers, chat interfaces, AI automation platforms, MCP gateways, and specialized AI tools.
Featured Templates
These flagship templates are highlighted in the catalog:
| Template | Description | Min RAM |
|---|---|---|
| OpenClaw | Open-source AI assistant platform with multi-model support | 2 GB |
| LibreChat | Multi-provider AI chat interface (OpenAI, Anthropic, local models) | 1 GB |
| Ollama | Local LLM inference server — run models on your hardware | 4 GB |
| Dify | LLM application development platform with visual workflow builder | 6 GB |
| DeerFlow | ByteDance super-agent harness — research, code, and create with a built-in sandbox (bring your own LLM key) | 4 GB |
Complete Template List
| Template | Description | Category | Min RAM | Security |
|---|---|---|---|---|
| OpenClaw | Open-source AI assistant platform | Chat | 2 GB | secure |
| LibreChat | Multi-provider AI chat interface | Chat | 1 GB | secure |
| Ollama | Local LLM inference server | Inference | 4 GB | advanced |
| Dify | LLM app development platform (model providers via plugins) | Platform | 6 GB | secure |
| Open WebUI | Web interface for Ollama and OpenAI-compatible APIs | Chat | 512 MB | secure |
| LiteLLM | OpenAI-compatible proxy for 100+ LLM providers | Proxy | 512 MB | secure |
| LocalAI | Self-hosted OpenAI-compatible API | Inference | 4 GB | advanced |
| Text Generation WebUI | Gradio web UI for large language models | Chat | 4 GB | advanced |
| Flowise | Drag-and-drop LLM flow builder | Platform | 1 GB | secure |
| Langflow | Visual framework for building LLM applications | Platform | 1 GB | secure |
| AnythingLLM | All-in-one AI desktop app turned self-hosted | Chat | 1 GB | secure |
| Jan | Open-source ChatGPT alternative, runs offline | Chat | 2 GB | secure |
| GPT4All | Local LLM chat with privacy focus | Chat | 4 GB | advanced |
| Serge | Self-hosted chat interface using llama.cpp | Chat | 2 GB | secure |
| SillyTavern | AI chat and roleplay frontend | Chat | 512 MB | secure |
| Stable Diffusion WebUI | Image generation with Stable Diffusion models | Image | 8 GB | advanced |
| ComfyUI | Node-based Stable Diffusion workflow editor | Image | 8 GB | advanced |
| Whisper | OpenAI Whisper speech-to-text server | Audio | 2 GB | secure |
| Coqui TTS | Deep learning text-to-speech server | Audio | 2 GB | secure |
| Haystack | LLM orchestration framework for RAG pipelines | Framework | 1 GB | secure |
| Chroma | AI-native open-source vector database | Database | 1 GB | secure |
| Qdrant | High-performance vector similarity search engine | Database | 1 GB | secure |
| Weaviate | Vector database for AI applications | Database | 2 GB | secure |
| ContextForge MCP Gateway | IBM MCP gateway & registry — federate MCP servers and tools behind one governed endpoint | Gateway | 1 GB | secure |
| DeerFlow | ByteDance super-agent harness (research, code, create) — bring your own LLM key | Agent | 4 GB | secure |
GPU Support
Templates marked with advanced security profile may support GPU passthrough. For GPU-accelerated inference (Ollama, LocalAI, Stable Diffusion), ensure your server has:
- NVIDIA GPU with CUDA support
nvidia-container-toolkitinstalled- Docker configured with the NVIDIA runtime
The template variables include GPU configuration options when applicable.
Connecting AI Services
Many AI services work well together. Common combinations:
- Ollama + Open WebUI — Local inference with a polished chat interface
- Ollama + LiteLLM — Local inference exposed as an OpenAI-compatible API
- Dify + Ollama — Visual AI workflow builder using local models
- LibreChat + LiteLLM — Multi-provider chat routing through a single proxy
- ContextForge + your MCP servers — Federate every MCP tool server behind one governed, authenticated MCP endpoint that clients like Claude Desktop or Cursor connect to
Services deployed on the same panel can communicate over the internal Docker network using their service names as hostnames.
ContextForge MCP Gateway — access and first login
ContextForge (IBM mcp-context-forge) is a single-container Model Context Protocol gateway and registry. It serves the browser Admin UI at /admin and the REST API + MCP endpoint on the same host and port (4444), reverse-proxied over HTTPS.
- First login: open the Open ContextForge Admin button (lands on
/admin) and sign in with the Admin Email you entered at deploy time plus the auto-generated Admin Password shown in the service's Credentials card (admin-only, audit-logged). On first sign-in ContextForge requires you to set a new password — that's its built-in password policy, not an error. Authentication is required on every request. - Secrets: the panel generates a unique JWT signing secret and the admin password per deployment — the JWT secret stays internal and is never shown or logged; only the admin password is surfaced.
- Persistence: registered servers, virtual tool catalogs, and config live in a SQLite database on a named volume (
/data), so they survive restarts and redeploys. The image runs as a non-root user, so the panel chowns the data volume on first boot before the gateway starts. - Health: the image is UBI-minimal (Python, no shell HTTP client), so the healthcheck is a Python
urllibprobe of/health; a crashed or hung container auto-recovers. - Image:
ghcr.io/ibm/mcp-context-forge, pinned to thev1.0.4release tag (amd64).
DeerFlow — access and first login
DeerFlow (ByteDance) is an open-source "super agent" harness that researches, codes, and creates — it plans long-horizon tasks, runs tools and shell commands in a built-in sandbox, and produces artifacts like reports, slides, and web pages. It deploys as a three-container stack — an nginx proxy (the only routed port, 2026), a Next.js frontend, and a FastAPI + LangGraph gateway — reverse-proxied over HTTPS. Only the nginx port is exposed; the frontend and gateway stay on the panel's private network.
- Bring your own LLM key (required): DeerFlow ships no built-in model, so the workspace is inert until you supply a key. Enter your provider key in the LLM API Key field at deploy time. It defaults to OpenAI; to use any other OpenAI-compatible provider (OpenRouter, DeepSeek, your own gateway), set the Provider Base URL (e.g.
https://openrouter.ai/api/v1) and a matching Model (e.g.google/gemini-2.5-flash). Optionally add a Tavily API Key to enable live web search. To change the key, model, or provider later, redeploy with the new values. - First login: open the Open DeerFlow button and create your admin account (email + password) on first visit — the workspace is locked until then, and there is no default password. DeerFlow then drops you into its chat workspace. This login is DeerFlow's own, separate from the panel.
- Secrets: the panel generates a unique frontend session secret and an internal gateway token per deployment — both stay internal and are never shown or logged. Your LLM key lives only in the gateway container's environment (encrypted at rest in the panel) and is never exposed to the browser.
- Sandbox: the agent runs tools and shell commands in an in-process sandbox — no Docker socket is mounted and no privileged container is created.
- Persistence: your chats, threads, and agent state live in a SQLite database on a named volume, so they survive restarts and redeploys.
- Health: the gateway image ships Python but no
curl/wget, so its healthcheck is a Pythonurllibprobe of/health; all three containers have healthchecks and start in dependency order (gateway → frontend → nginx). - Images:
ghcr.io/bytedance/deer-flow-backendandghcr.io/bytedance/deer-flow-frontend, pinned to thev2.0.0release tag (amd64 only).