Configuration

Server configuration and infrastructure setup for ISA MCP.

Overview

The platform uses config/init.yaml for service initialization and dependency configuration.

Service Configuration

MCP Server


mcp_service:
  port: 8081              # External port
  internal_port: 8300     # Internal port
  description: "Model Context Protocol service"
  depends_on:
    - postgres_grpc
    - qdrant
    - redis

Auto-Discovery


auto_discovery:
  enabled: true
  scan_interval_seconds: 300  # Re-scan every 5 minutes
  paths:
    tools: "tools/"
    prompts: "prompts/"
    resources: "resources/"
  patterns:
    tools: "*.py"
    prompts: "*.py"
    resources: "*.py"
  exclude:
    - "__pycache__"
    - "*.pyc"
    - "base_*.py"

Infrastructure Dependencies

PostgreSQL


postgres_grpc:
  port: 50061
  schemas:
    - mcp
  tables:
    - tool_executions    # Tool execution logs
    - prompt_templates   # Custom prompts
    - resource_cache     # Resource metadata

Qdrant (Vector Store)


qdrant:
  http: 6333
  grpc: 6334
  collections:
    - name: "mcp_tool_embeddings"
      vector_size: 1536
      distance: "Cosine"
    - name: "mcp_prompt_embeddings"
      vector_size: 1536
    - name: "mcp_document_embeddings"
      on_disk: true

Redis (Cache)


redis:
  port: 6379
  key_prefixes:
    - "mcp:tool:"       # Tool response cache
    - "mcp:prompt:"     # Prompt cache
    - "mcp:resource:"   # Resource cache
    - "mcp:embedding:"  # Embedding cache
    - "mcp:session:"    # Session state
    - "mcp:rate:"       # Rate limiting

MinIO (Object Storage)


minio:
  port: 9000
  buckets:
    - name: "mcp-resources"
      versioning: true
    - name: "mcp-cache"
      lifecycle:
        expiration_days: 30
    - name: "mcp-exports"
      lifecycle:
        expiration_days: 90

Startup Order


startup_order:
  - tier: 1
    description: "Core Infrastructure"
    services: [postgres, redis, qdrant, minio]
 
  - tier: 2
    description: "gRPC Gateways"
    services: [postgres_grpc, redis_grpc, qdrant_grpc]
 
  - tier: 3
    description: "MCP Service"
    services: [mcp]

Health Checks


health_checks:
  default_timeout_seconds: 5
  default_interval_seconds: 30
  default_failure_threshold: 3
 
  mcp_specific:
    tool_discovery:
      check: "count_registered_tools > 0"
    prompt_discovery:
      check: "count_registered_prompts > 0"
    resource_discovery:
      check: "count_registered_resources > 0"

Environment Variables

Variable	Description	Default
`MCP_PORT`	Server port	`8081`
`POSTGRES_HOST`	PostgreSQL host	`localhost`
`QDRANT_HOST`	Qdrant host	`localhost`
`REDIS_HOST`	Redis host	`localhost`
`LOG_LEVEL`	Logging level	`INFO`

Running the Server

Development


cd /path/to/isA_MCP
python main.py

With Custom Port


MCP_PORT=9000 python main.py

Docker


docker-compose up -d

Directory Structure


isA_MCP/
├── config/
│   └── init.yaml          # Main configuration
├── tools/                  # Auto-discovered tools
├── prompts/                # Auto-discovered prompts
├── resources/              # Auto-discovered resources
├── services/
│   └── skill_service/     # Skill management
├── core/
│   ├── auto_discovery.py  # Discovery logic
│   ├── security.py        # Security management
│   └── config.py          # Config loading
└── main.py                # Entry point

Adding External MCP Servers

See Server Aggregation for connecting to external MCP servers.


result = await client.call_tool("add_mcp_server", {
    "name": "external-server",
    "transport_type": "SSE",
    "connection_config": {
        "url": "https://external.example.com/sse"
    },
    "auto_connect": True
})

Connection Pooling

The SessionManager maintains a pool of active MCP server connections to avoid the overhead of reconnecting on every request.

How It Works

Lazy connect: connections are established on the first tool call, not at startup
LRU eviction: when the pool reaches capacity, the least-recently-used connection is evicted
Max 30 connections: configurable via MCP_POOL_MAX_SIZE env var


MCP_POOL_MAX_SIZE=30   # Maximum concurrent connections (default: 30)

Pool Stats

Monitor pool health:


curl http://localhost:8081/api/v1/mcp/pool/stats \
  -H "Authorization: Bearer YOUR_API_KEY"


{
  "active": 12,
  "idle": 4,
  "capacity": 30,
  "evictions": 3,
  "reconnections": 1
}

Field	Description
`active`	Connections currently handling requests
`idle`	Connections open but not active
`capacity`	Pool size limit
`evictions`	Total LRU evictions since startup
`reconnections`	Connections re-established after eviction

Server Priority Tiers

Assign a priority to each MCP server to control health check frequency and eviction behavior.


# Set priority when registering a server
{
  "name": "critical-tool-server",
  "url": "http://tool-server:8082",
  "priority": "CRITICAL"   # CRITICAL | STANDARD | OPTIONAL
}

Priority	Health check interval	LRU eviction
`CRITICAL`	Every 30 seconds	Exempt — never evicted
`STANDARD`	Every 60 seconds	Subject to eviction
`OPTIONAL`	Every 120 seconds	First to be evicted

Use CRITICAL for tool servers that are essential to your workflow (e.g., file system tools, code execution). Optional servers for rarely-used integrations can be OPTIONAL to free up pool capacity.

Regenerating Init Scripts


python -m scripts.generate_init

This generates initialization scripts based on config/init.yaml.

Next Steps

Quick Start - Getting started
Server Aggregation - External servers
Tools - Creating tools