Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Nosana MCP Agent list files in my project folder"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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.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-a5000Related MCP server: FPF Agent Stack
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
Bun 1.3+
Nosana CLI (npm install -g @nosana/cli)
Ollama (for local development)
Solana wallet + NOS tokens (for Nosana deployment)
# 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?"}'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 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"]
}
}
}
}
}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"]
}
}Type | Description | Required Fields |
| Local process via stdin/stdout |
|
| Remote server via HTTP SSE |
|
Method | Path | Description |
GET |
| Health check + uptime + model info |
GET |
| Agent info + MCP server list |
POST |
| Send message ( |
docker build -t nosana-mcp-agent .The image includes Ollama and will auto-pull the Qwen3 model on first startup.
docker run --gpus all -p 3000:3000 nosana-mcp-agentWithout GPU (CPU inference, much slower):
docker run -p 3000:3000 nosana-mcp-agentdocker tag nosana-mcp-agent SohniSwatantra/nosana-mcp-agent:latest
docker push SohniSwatantra/nosana-mcp-agent:latestEdit job-definition.json and replace YOUR_DOCKERHUB_USERNAME with your Docker Hub username.
# 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 \
--waitnosana market listnosana job get <job-address>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"
}
}
}
}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 templateVariable | Required | Default | Description |
| No |
| Model for Ollama to pull and serve |
| No |
| Embedding model |
| No |
| Ollama server bind address |
| No |
| HTTP server port |
| No |
| Enable MCP stdio server mode |
| No | — | Set to |
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 |
This server cannot be deployed
Native Solana staking for AI agents. 26 MCP tools, one-shot signing, webhooks.
HiveCompute MCP Server — decentralized inference router for AI agents
AI agent registry — search, discover, register, and connect agents via MCP.
Open-source Zapier/n8n alternative as an MCP server: agents build, run and debug your workflows.
