BangC MCP Server

bangc-mcp-server is a self-contained MCP server for searching embedded BangC, CNRT, and CNNL reference material. It ships prebuilt Rust binaries and exposes a small Python launcher so the server can be started with uvx.

uvx bangc-mcp-server

HTTP mode is also available:

uvx bangc-mcp-server serve --http --addr 127.0.0.1:8765

The command line is organized around subcommands:

uvx bangc-mcp-server serve
uvx bangc-mcp-server serve --http --addr 127.0.0.1:8765
uvx bangc-mcp-server stats

Generate a clangd-compatible compilation database for .mlu files in the current project:

uvx bangc-mcp-server ccdb -r . -a mtp_372

The default output is the standard compile_commands.json file. Use -o compile_command.json only for tools that require a custom filename. The command also updates .vscode/settings.json with clangd arguments for --query-driver and --compile-commands-dir, and writes a small clangd-only shim under target/compile_commands/ so BANG built-in register helpers such as taskId and taskDimX resolve during indexing.

Compile a single .mlu file from the CLI:

uvx bangc-mcp-server compile-object kernel.mlu -a mtp_372
uvx bangc-mcp-server compile-extension kernel.mlu -a mtp_372 --smoke-import

MCP Configuration

Use the uvx command as the MCP server command:

{
  "mcpServers": {
    "bangc-docs": {
      "command": "uvx",
      "args": ["bangc-mcp-server"]
    }
  }
}

Included Tools

• bangc_search
• bangc_api_lookup
• bangc_symbols
• bangc_get_chunk
• bangc_get_doc
• bangc_tutorials
• bangc_context_pack
• bangc_list_documents
• bangc_list_sources
• neuware_api_search
• neuware_api_lookup
• neuware_env_check
• mlu_device_query
• python_torch_env_check
• bangc_compile_torch_extension
• bangc_compile_torch_extension_plan

The mlu_device_query tool uses bounded one-shot cnmon commands when the local Neuware runtime is available. Mutating or continuously refreshing monitor modes are intentionally not exposed.

The bangc_compile_torch_extension tool compiles a local .mlu file into a Python torch extension for debugging on a machine with Neuware, cncc, PyTorch, and torch_mlu installed. It generates a small pybind11 binding around the wrapper function and does not run task-specific reference checks. Successful compile results are intentionally compact; use bangc_compile_torch_extension_plan to inspect the full environment and command lines.

Supported Platforms

This package currently includes:

• Windows x86_64
• Linux x86_64, musl-linked

Set BANGC_MCP_SERVER_BIN=/path/to/bangc-mcp-server to override the bundled binary.

Notes

The package README intentionally avoids local filesystem paths, private environment output, usernames, hostnames, and device-specific command output.