How to Use Tool Filtering to Limit Exposed MCP Tools in code-review-graph

You can restrict which MCP tools are available in code-review-graph by using the --tools CLI flag or the CRG_TOOLS environment variable to specify a comma-separated allow-list of tool names.

code-review-graph ships with 30+ MCP tools by default, which can overwhelm LLM token limits or introduce latency in resource-constrained environments. The repository provides a built-in filtering mechanism that lets you expose only the specific tools your workflow requires, without modifying any source code. This capability is implemented in code_review_graph/main.py and documented in the README under the Tool Filtering section.

How Tool Filtering Works

The filtering logic centers on the helper function _apply_tool_filter in code_review_graph/main.py. This function receives the full set of available tools and removes any tool whose name is not present in your supplied allow-list.

According to the source code at lines 22-48, _apply_tool_filter compares each tool name against the filtered list before the server starts. When the server initializes via main(), it invokes this filter before the MCP event loop begins (lines 85-92), ensuring that only your specified tools are registered with FastMCP.

Configuration Methods

You can supply the tool allow-list through two mechanisms, with the CLI flag taking precedence over the environment variable. If neither is set, all 30+ tools remain exposed.

CLI Flag (Preferred Method)

Pass a comma-separated list of tool names to the --tools argument when starting the server:

code-review-graph serve \
    --tools query_graph_tool,semantic_search_nodes_tool,detect_changes_tool

This method provides the highest precedence and is recommended for per-project configurations.

Environment Variable (Fallback)

Set the CRG_TOOLS environment variable before launching the server:

export CRG_TOOLS=query_graph_tool,semantic_search_nodes_tool
code-review-graph serve

This approach is useful for containerized deployments or shared development environments where you want consistent defaults.

Client Configuration Example

When configuring the server in a client like Claude Desktop or Cursor, include the tools flag in the JSON configuration:

{
  "mcpServers": {
    "code-review-graph": {
      "command": "code-review-graph",
      "args": [
        "serve",
        "--tools",
        "query_graph_tool,semantic_search_nodes_tool,detect_changes_tool,get_review_context_tool"
      ]
    }
  }
}

The CLI argument parsing is handled in code_review_graph/cli.py, which forwards the values to the main entry point.

Key Implementation Details

The tool filtering feature involves three core components:

• code_review_graph/main.py – Contains _apply_tool_filter and the orchestration logic that applies filtering before server startup.
• code_review_graph/cli.py – Handles argument parsing for the --tools flag.
• README.md – Documents the filtering behavior and precedence rules (lines 12-24).

Summary

• Use the --tools CLI flag or CRG_TOOLS environment variable to limit exposed MCP tools.
• The CLI flag takes precedence over the environment variable when both are present.
• Filtering occurs in _apply_tool_filter before the MCP event loop starts, ensuring only requested tools are registered.
• All 30+ tools remain available if no filter is specified.

Frequently Asked Questions

What happens if I specify a tool name that doesn't exist?

The server ignores invalid tool names silently during the filtering process. Only tools that exist in the default set and match your allow-list are registered; non-existent names are simply skipped without raising an error.

Can I use wildcards or patterns for tool names?

No, the filtering mechanism requires explicit tool names. As implemented in _apply_tool_filter, the comparison is a direct string match against the allow-list, so you must specify each tool name exactly as defined in the source code.

Does tool filtering affect server startup time?

Startup time remains negligible because _apply_tool_filter performs a simple list comparison on the existing tool registry. The operation occurs before the FastMCP event loop begins, adding no perceptible latency to the initialization process.

How do I see the full list of available tool names?

Check the repository documentation or examine the tool registration in code_review_graph/main.py. The README (lines 12-24) provides examples of commonly used tool names, and the source code defines all 30+ available tools that you can include in your filter.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}