Agent Platform

Eine schlanke, business-taugliche Agent-Plattform mit LiteLLM, Pydantic AI und MCP.

Architektur

┌─────────────────────────────────────────────┐
│  agent-platform (Port 8000)                 │
│  - FastAPI + LiteLLM + Pydantic AI          │
│  - HTMX-UI (kein JS-Build)                  │
│  - SQLite                                   │
│  - Audit-Log, Auth                          │
└──────────────────┬──────────────────────────┘
                   │ MCP (HTTP)
                   ▼
┌─────────────────────────────────────────────┐
│  mcp-tools (Port 8501, intern)              │
│  - fastmcp                                  │
│  - 6 generische Tools                       │
└─────────────────────────────────────────────┘

Quickstart

# 1. Env-Datei anlegen
cp .env.example .env
# .env editieren: LLM_API_KEY und AUTH_TOKEN setzen

# 2. Starten
docker compose up -d --build

# 3. UI öffnen
http://localhost:8000

# 4. Ersten Agent anlegen
# Im UI: Agents → "Neuen Agent erstellen"
# System-Prompt: "Du bist ein hilfreicher Assistent."
# Erlaubte Tools: "echo, calculate, get_time"

API

# Health
curl http://localhost:8000/health

# Agents
curl -H "Authorization: Bearer $AUTH_TOKEN" http://localhost:8000/api/agents

# Chat
curl -X POST -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"message": "Hallo!"}' \
  http://localhost:8000/api/chat/general_assistant

# MCP-Tools
curl http://localhost:8000/api/tools

Verfügbare MCP-Tools

• echo(text) - Echo-Tool
• get_time(timezone_name) - Aktuelle Zeit
• calculate(expression) - Mathe-Ausdruck (sicher)
• http_get(url) - HTTP-Request (SSRF-Schutz)
• list_env(prefix) - Umgebungsvariablen
• json_format(data) - JSON formatieren

Konfiguration

Alle Env-Vars sind in .env.example dokumentiert.

Entwicklung

# Backend lokal starten (ohne Docker)
cd agent_platform
pip install -e .
LLM_API_KEY=sk-xxx uvicorn api:app --reload

# MCP-Server lokal
cd mcp_tools
pip install -r requirements.txt
python server.py

Ressourcen

• agent-platform: 256 MB RAM, 0.5 CPU
• mcp-tools: 128 MB RAM, 0.25 CPU
• SQLite: 5-10 MB

Deploy (Coolify)

1. In Coolify: Projects → New Project → name it (e.g. agent-platform).
2. Inside the project: Resources → New → Docker Compose.
3. Repository URL: https://forgejo.media-on.de/Leopoldadmin/agent-platform.git, Branch: main.
4. Domain: bind a hostname to the agent-platform service (Traefik routes via the expose: 8000 label).
5. Environment Variables: set AUTH_TOKEN (≥32 chars, NOT change-me-in-production) and LLM_API_KEY; other vars can keep their .env.example defaults.
6. Click Deploy — Coolify builds the two images, creates the agent-data volume, and brings up the stack.
7. Health check: curl https://<your-domain>/health returns {"status":"ok"} once mcp-tools is healthy and agent-platform is up.