ollama-almasrv-mcp

MCP (Model Context Protocol) server for accessing Ollama models on a remote GPU server.

Designed for ALMASRV (NVIDIA RTX PRO 6000 96GB VRAM) but works with any Ollama instance behind a compatible HTTP gateway.

Features

• 6 MCP tools: chat, think, embed, similarity, models, health
• 10 models: 4 local (llama3.3:70b, qwen3:32b, llama3.2:3b, mxbai-embed-large) + 6 cloud
• 1024-dim embeddings compatible with SQL Server 2025 VECTOR(1024)
• Thinking models with reasoning traces (qwen3:32b, kimi-k2-thinking, glm-5, kimi-k2.5, minimax-m2.5)
• Configurable endpoints via environment variables

Quick Install

pip install ollama-almasrv-mcp

Setup with Claude Code

# Add to Claude Code (user-level, available everywhere)
claude mcp add --scope user ollama-almasrv -- ollama-almasrv-mcp

# Or with custom server URLs
claude mcp add --scope user \
  -e ALMASRV_GATEWAY_URL=http://your-server:8030 \
  -e ALMASRV_EMBED_URL=http://your-server:8031 \
  ollama-almasrv -- ollama-almasrv-mcp

Available Tools

Tool	Description
ollama_chat	Chat with any Ollama model (default: qwen3:32b)
ollama_think	Chat with thinking models that return reasoning traces
ollama_embed	Generate 1024-dim embedding vectors
ollama_similarity	Calculate cosine similarity between two texts
ollama_models	List all available models
ollama_health	Check gateway and embedding service health

Configuration

Environment Variable	Default	Description
ALMASRV_GATEWAY_URL	http://192.168.50.78:8030	Ollama gateway (chat/think/models)
ALMASRV_EMBED_URL	http://192.168.50.78:8031	Embedding service (embed/similarity)

Requirements

• Python >= 3.10
• A running Ollama instance with a compatible HTTP gateway
• Gateway endpoints: /chat, /models, /health
• Embedding endpoints: /embeddings, /similarity, /health

License

MIT