# MCP-Ollama Client

Small command‑line chat client that

1. **Runs entirely offline with a local LLM via Ollama**
2. **Talks to any number of Model‑Context‑Protocol (MCP) servers**, all declared in one `config.json`

At start‑up the client launches every server, fetches their tool schemas, prefixes tool names (`postgres.*`, `filesystem.*`, …) and gives the merged list to the model.
The LLM decides which server to call for each user request.

---

## Features

| Feature                         | Notes                                                                                                               |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------------- |
| **Local LLM first**             | Default model is `qwen3:14b`, but any function‑calling model that Ollama exposes will work. No cloud keys required. |
| **Multi‑server out‑of‑the‑box** | Postgres, filesystem, or your own MCP servers can run side‑by‑side; all are defined in `config.json`.               |
| **Collision‑free tool names**   | Tools are exposed as `<server>.<tool>` so names never clash.                                                        |

---

## Requirements

| Component     | Version tested                             |
| ------------- | -------------------------------------------|
| Python        | ≥ 3.12                                     |
| Ollama        | ≥ 0.8.0                                    |
| MCP server(s) | Anything that supports **stdio** transport |

---

## Quick start

```bash
# 1. clone
git clone https://github.com/Nagharjun17/MCP-Ollama-Client.git
cd mcp-ollama-client

# 2. set up environment
uv venv
source .venv/bin/activate
uv pip sync pyproject.toml         # or: uv pip install -r uv.lock

# 3. pull a local model
ollama pull qwen3:14b

# 4. edit Model Name, DATABASE_URI etc. in config.json and download required MCP Servers. In my case:
uv pip install postgres-mcp

# 5. run
uv run client.py
```

Example boot log:

```
🔌 Starting MCP server: postgres
🔌 Starting MCP server: filesystem
🛠️  Aggregated tools: ['postgres.list_schemas', 'filesystem.read_file', ...]
>>>
```

![MCP Client Startup](assets/start-mcp-client.png)

Type natural language queries; the model will decide when and how to call the exposed tools.

---

## `config.json` format

```jsonc
{
  "llm": {
    "model": "qwen3:14b",
    "temperature": 0.7,
    "max_tokens": 2048
  },

  "mcpServers": {
    "postgres": {
      "command": "postgres-mcp",
      "args": ["--access-mode=restricted"],
      "env": {
        "DATABASE_URI": "postgresql://user:pass@localhost:5432/db"
      }
    },

    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/mnt/smbshare/"
      ]
    }
  }
}
```

*Keys under **`mcpServers`** become prefixes (**`postgres.*`**, **`filesystem.*`**).*

*Each server starts as its own **`stdio`** subprocess.*

*Add or remove servers without touching **`client.py`**.*

Everything stays local, everything is configurable in one file.