Nabu

You run a pair of MCP servers that give a code intelligence workspace to your agent. One server provides semantic code inspection and a structured code view, while the other exposes a workspace-style interface that the agent can use as a TUI-backed context. Together, they help you search, understand, and navigate large codebases more efficiently and with lower token usage.

How to use

Launch the MCP servers and connect your MCP client to them to start using semantic code search, structural views, and workspace organization. You will use two servers: an index/engine that runs locally and a proxy-based HTTP server that the agent communicates with.

1. Start the local code intelligence engine and expose an MCP endpoint. This server runs as a local stdio process and is invoked with a Python module plus a rich codebase configuration. It provides HTTP access for other components and manages the internal index, codeframes, and search capabilities.

2. Start the workspace TUI proxy for Claude Code interaction. This is an HTTP MCP endpoint that the agent talks to for workspace management and augmented context. The proxy coordinates with the code intelligence engine to keep the agent’s view synchronized with the current project state.

3. Run the Claude Code client and point it at the proxy so the agent can issue requests, inspect code, and use the workspace UI seamlessly. The agent will use the connected MCP servers to fetch symbols, navigate the code structure, and perform semantic searches across the codebase.

How to install

Prerequisites you need on your machine:

• Python 3.13 (tested; may work with >= 3.11)

• Git for repository operations

• Optionally, a virtual environment (recommended)

Steps to install and run the MCP setup:

1) Install the code intelligence prototypes

Clone the repository and install the MCP package in editable mode to enable local development and testing.

$ git clone https://github.com/y3i12/nabu_nisaba.git
$ cd nabu_nisaba
$ pip install -e .

2) Configure the MCP servers

Create or edit the MCP configuration file to define the two servers. The following configuration is used to start the two MCP servers as described.

{
  "mcpServers": {
    "nabu": {
      "command": "python",
      "args": [
        "-m",            "nabu.mcp.server",
        "--codebase",    "nabu:/path/to_codebase/:/path/to/database.kuzu:active:true",
        "--context",     "development",
        "--enable-http",
        "--http-port",   "1338",
        "--dev-mode"
      ],
      "env": {
        "PYTHONPATH": "/path/to/nabu_nisaba/",
        "NABU_LOG_LEVEL": "INFO",
        "NABU_MODEL_CACHE": "/path/to/nabu_nisaba/.nabu/hf_cache/"
      }
    },
    "nisaba": {
      "type": "http",
      "url": "http://localhost:9973/mcp"
    }
  }
}

3) Start Claude Code with the proxy

Start the Claude Code proxy-wrapped workflow that connects to the MCP servers.

$ python -m nisaba.cli claude [--continue]

4) Optional: configure workspace status and precompact hooks

If you want a compact workspace status line and precompact hooks, update the Claude Code settings with a status line command and the precompact script invocations.

{
  "statusLine": {
    "type": "command",
    "command": "./scripts/workspace-status",
    "padding": 0
  },
  "hooks": {
    "PreCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python3 ./scripts/precompact_extract.py"
          }
        ]
      }
    ]
  },
  "alwaysThinkingEnabled": true
}

Notes

Always think between tool execution and message exchange is strongly recommended as it helps the agent maintain a thoughtful workflow and reduces context burst.