Nosana MCP Agent

ElizaOS AI agent + MCP tools + Qwen3.5 9B on decentralized GPU

An open-source AI agent powered by Qwen3.5 (9B) running locally via Ollama on Nosana's decentralized GPU network. No external API keys needed for inference. Connects to external tools via MCP (Model Context Protocol).

Fork it. Configure it. Deploy it to a GPU in one command.

TL;DR — Deploy in 3 Commands

git clone https://github.com/SohniSwatantra/nosana-mcp-agent.git && cd nosana-mcp-agent
make push DOCKER_USER=your-dockerhub-username
make deploy NOSANA_MARKET=nvidia-a5000

What This Does

• Local LLM: Qwen3.5 9B running on GPU via Ollama — no API keys needed
• MCP Client: Connects to external MCP servers (filesystem, GitHub, etc.) to access tools
• MCP Server: Exposes the agent as an MCP server for Claude Desktop and other MCP clients
• HTTP API: REST endpoints on port 3000 for health checks, info, and chat
• GPU-Optimized: Containerized with Ollama for Nosana GPU deployment
• Open Source: MIT licensed, fork and customize

Prerequisites

• Bun 1.3+
• Docker
• Nosana CLI (npm install -g @nosana/cli)
• Ollama (for local development)
• Solana wallet + NOS tokens (for Nosana deployment)

Quick Start (Local)

# 1. Install Ollama and pull the model
ollama pull qwen3.5:9b
ollama pull nomic-embed-text:latest

# 2. Install dependencies
bun install

# 3. Start the agent (Ollama must be running)
bun run start

# 4. Test it
curl http://localhost:3000/health
curl -X POST http://localhost:3000/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Hello, what tools do you have?"}'

Model Configuration

The agent uses Qwen3.5 9B by default, configured in character.json:

{
  "settings": {
    "OLLAMA_URL": "http://localhost:11434",
    "OLLAMA_SMALL_MODEL": "qwen3.5:9b",
    "OLLAMA_LARGE_MODEL": "qwen3.5:9b",
    "OLLAMA_EMBEDDING_MODEL": "nomic-embed-text:latest"
  }
}

To use a different model, change the model names in character.json and the OLLAMA_MODEL env var in the Dockerfile/job definition. Qwen3.5 9B is 6.6GB and needs ~8GB VRAM — fits comfortably on an RTX A5000 (16GB) or RTX 5000.

MCP Configuration

MCP server connections are configured in character.json under settings.mcp.servers:

{
  "settings": {
    "mcp": {
      "servers": {
        "filesystem": {
          "type": "stdio",
          "command": "npx",
          "args": ["-y", "@modelcontextprotocol/server-filesystem", "/app/data"]
        }
      }
    }
  }
}

Adding More MCP Servers

Edit character.json to add servers:

{
  "github": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "your-token" }
  },
  "puppeteer": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-puppeteer"]
  }
}

Supported Server Types

Type	Description	Required Fields
stdio	Local process via stdin/stdout	command, args
sse	Remote server via HTTP SSE	url

HTTP API Endpoints

Method	Path	Description
GET	/ or /health	Health check + uptime + model info
GET	/info	Agent info + MCP server list
POST	/chat	Send message ({"message": "...", "userId?": "..."})

Docker

Build

docker build -t nosana-mcp-agent .

The image includes Ollama and will auto-pull the Qwen3 model on first startup.

Run Locally (requires NVIDIA GPU + Docker GPU support)

docker run --gpus all -p 3000:3000 nosana-mcp-agent

Without GPU (CPU inference, much slower):

docker run -p 3000:3000 nosana-mcp-agent

Push to Docker Hub

docker tag nosana-mcp-agent SohniSwatantra/nosana-mcp-agent:latest
docker push SohniSwatantra/nosana-mcp-agent:latest

Deploy to Nosana

1. Update job-definition.json

Edit job-definition.json and replace YOUR_DOCKERHUB_USERNAME with your Docker Hub username.

2. Post the Job

# Deploy to RTX A5000 market (16GB VRAM, ideal for Qwen3.5 9B)
nosana job post \
  --file job-definition.json \
  --market nvidia-a5000 \
  --gpu \
  --wait

# Or target RTX 4090 (24GB VRAM)
nosana job post \
  --file job-definition.json \
  --market nvidia-4090 \
  --gpu \
  --wait

3. Check Available GPU Markets

nosana market list

4. Monitor Your Job

nosana job get <job-address>

Claude Desktop Integration

To use this agent as an MCP server from Claude Desktop (requires local Ollama):

{
  "mcpServers": {
    "nosana-agent": {
      "command": "bun",
      "args": ["run", "start"],
      "cwd": "/path/to/nosana-mcp-agent",
      "env": {
        "MCP_STDIO": "true"
      }
    }
  }
}

Project Structure

nosana-mcp-agent/
  character.json       # Agent character + model + MCP server configuration
  server.ts            # Main agent server (HTTP + MCP)
  entrypoint.sh        # Docker entrypoint (starts Ollama, pulls model, starts agent)
  test-client.ts       # HTTP API test client
  Dockerfile           # Production container with Ollama
  job-definition.json  # Nosana GPU deployment definition
  package.json         # Dependencies and scripts
  data/                # Directory accessible to MCP filesystem server
  .env.example         # Environment variable template

Environment Variables

Variable	Required	Default	Description
OLLAMA_MODEL	No	qwen3.5:9b	Model for Ollama to pull and serve
OLLAMA_EMBEDDING_MODEL	No	nomic-embed-text:latest	Embedding model
OLLAMA_HOST	No	0.0.0.0:11434	Ollama server bind address
PORT	No	3000	HTTP server port
MCP_STDIO	No	false	Enable MCP stdio server mode
NODE_ENV	No	—	Set to production in Docker

GPU Requirements

Model	Size	VRAM Required	Recommended Nosana Market
qwen3.5:4b	2.7GB	~4GB	nvidia-a4000, nvidia-3060-community
qwen3.5:9b	6.6GB	~8GB	nvidia-a5000, nvidia-4090
qwen3.5:14b	9.5GB	~12GB	nvidia-a5000, nvidia-4090
qwen3.5:32b	21GB	~24GB	nvidia-a100-40gb, nvidia-6000-ada