nuntius-mcp

nuntius (Latin) — messenger, herald; one who carries word between those who cannot speak directly.

nuntius-mcp is an MCP server for Claude Code that delegates agentic tasks to local or cloud Ollama models. Claude invokes a single run_agent tool; the Ollama model runs autonomously in a tool-use loop with access to all MCP tools configured in Claude Code — web search, vector memory, browser automation, file I/O, bash, and more. When it's done, the result comes back to Claude as a string.

Claude offloads the heavy lifting. The local model does the work.

How It Works

Claude Code  ──run_agent──▶  nuntius-mcp (MCP server)
                                  │
                                  ├─ reads ~/.claude.json + plugin caches
                                  ├─ discovers all configured MCP tools
                                  └─▶  Ollama model (agentic loop)
                                            │
                                   ┌────────┴────────┐
                                   │  tool call?      │
                                   │  yes → execute   │
                                   │  no  → return ◀──┘
                                            │
Claude Code  ◀──── final answer ────────────┘

On each run_agent call, nuntius-mcp reads your Claude Code MCP configuration, proxies every enabled tool to the Ollama model under a server__tool namespace (e.g. searxng__search, playwright__browser_screenshot), then runs the model in a tool-use loop until it produces a final answer or hits the iteration limit.

Features

• Full MCP proxy — automatically surfaces every MCP tool Claude Code has configured; no manual wiring
• Server namespacing — tools are named server__tool so there are never collisions
• Opt-in or opt-out — include/exclude specific MCP servers per call or globally in config
• Model aliases — short names (gpt, fast, etc.) resolve to full Ollama model strings
• Built-in fallbacks — file I/O and bash tools when no MCP server covers the capability
• Run logging — every completed run appended to logs/runs.jsonl
• Dual-mode — same package runs as an MCP stdio server or a web UI config manager
• Optional web UI — dashboard, model aliases, tool defaults, and endpoint settings

Requirements

• Python 3.11+
• Ollama running and reachable
• Claude Code (the MCP server is invoked as a subprocess by Claude Code)
• Optional: SearXNG for web search, Qdrant for vector memory

Installation

# Development install (recommended)
uv pip install -e ".[dev]"

# Or with pip
pip install -e ".[dev]"

Configuration

Edit config/config.toml. The file is hot-reloaded on each run_agent call — no restart needed after UI changes.

[server]
log_level = "info"
max_iterations = 20

[ollama]
base_url = "http://your-ollama-host:11434"

[searxng]
base_url = "http://your-searxng-host:8888"

[ui]
host = "127.0.0.1"   # set "0.0.0.0" to expose on LAN (no auth implemented)
port = 7860

[models]
default = "gpt"      # alias used when model="" in run_agent

[models.aliases]
gpt  = "gpt-oss:120b"
glm  = "glm-5:cloud"
kimi = "kimi-k2.5:cloud"
max  = "minimax-m2.7:cloud"
qwen = "qwen3-coder:480b-cloud"

[tools]
enable_all = false   # set true to enable all tools for every call
allowed = []         # pre-enable specific tools globally, e.g. ["searxng__search"]

[mcp]
# exclude = ["some-server"]   # opt-out specific servers from being proxied
# include = ["searxng"]       # uncomment to switch to opt-in mode (only named servers)

Claude Code Integration

Add to ~/.claude/settings.json under mcpServers:

{
  "mcpServers": {
    "nuntius-mcp": {
      "command": "python",
      "args": ["-m", "src"],
      "cwd": "/path/to/nuntius-mcp",
      "env": { "NUNTIUS_MODE": "mcp" }
    }
  }
}

Or with uv:

{
  "mcpServers": {
    "nuntius-mcp": {
      "command": "uv",
      "args": ["run", "python", "-m", "src"],
      "cwd": "/path/to/nuntius-mcp"
    }
  }
}

After saving, restart Claude Code. The run_agent tool will appear in Claude's tool list.

run_agent Parameters

Example — web research with a local model:

run_agent(
  prompt="Summarise the three most recent papers on speculative decoding",
  model="smart",
  tools=["searxng"]
)

Example — file task with bash:

run_agent(
  prompt="Find all TODO comments in src/ and write a summary to docs/todos.md",
  enable_all=true
)

Example — Ollama cloud model:

Ollama supports cloud-hosted models via the :cloud tag. These work identically to local models — just pass the full model name or a configured alias:

run_agent(
  prompt="Review this codebase for security issues and write a findings report",
  model="kimi-k2.5:cloud",
  tools=["searxng", "qdrant"],
  system_prompt="You are a security-focused code reviewer. Be thorough and cite line numbers."
)

MCP Tool Proxy

At startup, nuntius-mcp reads MCP server configurations from:

1. ~/.claude.json — global and project-level servers synced by Claude Code
2. ~/.claude/plugins/cache/**/.mcp.json — plugin-installed servers

Each server is connected via stdio and its tools are listed. Tools are registered as server__tool_name (double underscore). The ollama-mcp and nuntius-mcp servers are always excluded to prevent recursion.

Scoping tools per call:

# Enable all tools from the searxng server
tools=["searxng"]

# Enable specific tools by full name
tools=["searxng__search", "qdrant__search"]

# Enable everything
enable_all=true

Note: Large tool lists (40–60 tools across many servers) can degrade smaller Ollama models. Use tools=["server_name"] to scope what the model sees.

Optional Web UI

Start the UI server alongside the MCP server:

NUNTIUS_MODE=ui python -m src
# Serves on http://127.0.0.1:7860

The MCP server itself has zero dependency on the UI — it runs fine without it.

Config changes via the UI write directly to config/config.toml and take effect on the next run_agent call.

Docker

# MCP server only
docker compose up nuntius-mcp -d

# MCP server + web UI
docker compose --profile ui up -d

Important: Claude Code invokes the MCP server as a local subprocess via stdio. The Docker container cannot receive that stdio connection and is not used for normal operation. It is provided for integration testing and self-hosted UI deployments.

network_mode: "host" is used so the container can reach LAN services (Ollama, SearXNG, etc.). This requires Linux — it does not work on Docker Desktop (Mac/Windows).

Copy .env.example to .env and set APP_UID / APP_GID to match your user before starting.

Development

# Run all tests
pytest tests/ -v

# Run a single test
pytest tests/test_agent.py::test_agent_no_tools_returns_content -v

# Inspect the MCP tool list interactively
mcp dev src/server.py

Project structure:

nuntius-mcp/
├── config/
│   └── config.toml          # runtime config (edit or use UI)
├── logs/
│   └── runs.jsonl           # one JSON line per agent run
├── src/
│   ├── __main__.py          # entrypoint — routes to mcp or ui mode
│   ├── server.py            # FastMCP stdio server + run_agent tool
│   ├── agent.py             # Ollama agentic loop
│   ├── config.py            # config loading + model alias resolution
│   └── tools/
│       ├── registry.py      # tool registry + enable/disable logic
│       ├── mcp_proxy.py     # MCP server discovery + tool forwarding
│       ├── file_ops.py      # read_file, write_file, edit_file, glob_files, grep_content
│       └── bash.py          # bash (off by default)
├── tests/
├── docker-compose.yml
├── Dockerfile
└── pyproject.toml