Mnevis Server

⚠️ This is an experiment.

A lightweight, zero-dependency Python MCP server that exposes a single do_everything tool. Any AI agent that supports MCP can use it to offload all language-model work to a local OpenAI-compatible endpoint, reducing cost on the agent's primary LLM.

---

How it works

AI Agent (e.g. Bob/Claude/Copilot/Cursor)
    │
    │  MCP stdio (JSON-RPC 2.0)
    ▼
mnevis  server.py
    │
    │  HTTP POST /v1/chat/completions
    ▼
Local LLM  (Ollama, LM Studio, llama.cpp, vLLM, …)

The agent calls the do_everything tool with a prompt (and optional system instruction).
The server forwards the request to the local LLM using the standard OpenAI chat-completions API
and returns the model's response to the agent.

The tool description is worded so that any LLM automatically understands it should delegate
every task to the tool instead of reasoning on its own.

---

Requirements

• Python 3.11+
• No third-party packages — uses the standard library only (urllib, json, sys, os)
• A running local LLM that exposes a /v1/chat/completions endpoint
  (e.g. Ollama, LM Studio, llama.cpp server, vLLM)

---

Configuration

All settings are read from environment variables at startup:

Variable	Default	Description
MNEVIS_URL	http://localhost	Base URL of the local LLM server
MNEVIS_PORT	11434	Port the LLM server listens on
MNEVIS_MODEL	llama3	Model name to pass in the request
MNEVIS_API_KEY	(empty)	Optional API key (sent as Bearer token)

Examples

Ollama (default port 11434):

MNEVIS_MODEL=llama3 python server.py

LM Studio (default port 1234):

MNEVIS_URL=http://localhost MNEVIS_PORT=1234 MNEVIS_MODEL=lmstudio-community/Meta-Llama-3-8B-Instruct python server.py

vLLM with API key:

MNEVIS_URL=http://my-gpu-box MNEVIS_PORT=8000 MNEVIS_MODEL=mistral-7b MNEVIS_API_KEY=secret python server.py

---

Running the server

The server communicates over stdio (JSON-RPC 2.0), so it is spawned as a child process by
the MCP host — you do not run it manually in most cases.

To test it directly:

python server.py

Then paste a raw JSON-RPC message, e.g.:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.0.1"}}}

---

Registering with an MCP host

Bob / Cursor / Claude Desktop

Add to your mcp.json (workspace or global):

{
  "mcpServers": {
    "mnevis": {
      "command": "python",
      "args": ["/absolute/path/to/mnevis-mcp/server.py"],
      "env": {
        "MNEVIS_URL":   "http://localhost",
        "MNEVIS_PORT":  "11434",
        "MNEVIS_MODEL": "llama3",
        "MNEVIS_API_KEY": ""
      }
    }
  }
}

Replace the args path with the actual absolute path on your machine.
Set LOLA_PORT / LOLA_MODEL to match your local LLM setup.

---

Exposed tool

do_everything

Argument	Type	Required	Description
prompt	string	✅	The full task, question, or conversation to process
system	string	❌	Optional system / persona instruction for the local LLM

The tool description explicitly instructs the calling agent to send every task here rather
than reasoning itself, ensuring maximum cost offloading.

---

Project layout

mnevis-mcp/
├── server.py        # MCP server (single file, stdlib only)
├── pyproject.toml   # Project metadata
├── README.md        # This file
└── .gitignore

---

License

MIT