使用说明
ctroy-code-analysis
An MCP server that provides six code review tools for use with Claude Code (or any MCP client). Each reviewer reads a file from disk, pairs its contents with a structured review prompt, and returns the bundle for the LLM to evaluate.
Installation
pip install ctroy-code-analysis
Requires Python 3.10+.
Connecting to Claude Code
Add the server to your project's .mcp.json (per-project) or
~/.claude.json (global):
{
"mcpServers": {
"ctroy-code-analysis": {
"command": "ctroy-code-analysis"
}
}
}
Restart Claude Code. The six tools will appear automatically.
Tools
Each tool takes a filepath (or directory), reads the contents, and returns them alongside review instructions. The LLM then generates the review.
| Tool | Input | What It Reviews |
|---|---|---|
review_comments |
filepath |
Identifies superfluous comments that restate obvious code and inaccurate comments that contradict what the code does. |
review_names |
filepath |
Identifies unclear, inaccurate, or shadowed names in classes, functions, variables, and constants. |
review_cohesion |
filepath |
Identifies related logic scattered across the codebase that should be colocated. |
review_performance |
filepath |
Identifies unnecessary computation, inefficient algorithms, and missed optimization opportunities. |
review_test_coverage |
filepath |
Identifies untested code paths including edge cases, error paths, and boundary conditions. |
draw_gridmat |
directory |
Lists the directory structure and produces an ASCII execution-path diagram with emoji-coded entry points. |
Example usage in Claude Code
Ask Claude naturally:
Use review_comments on src/parser.py
Use review_performance on lib/data_pipeline.py
Use draw_gridmat on the src/ directory
Prompts
Each tool also has a corresponding prompt (prefixed with prompt_). Prompts
return just the review instructions without reading any files, so the LLM
applies them to code it already has in context.
| Prompt | Input |
|---|---|
prompt_review_comments |
filepath |
prompt_review_names |
filepath |
prompt_review_cohesion |
filepath |
prompt_review_performance |
filepath |
prompt_review_test_coverage |
filepath |
prompt_draw_gridmat |
directory |
What the reviewers look for
Comment review flags two categories, ordered by severity: inaccurate comments (say something the code doesn't do) and superfluous comments (restate what the code clearly says).
Name review flags five categories: unclear names, inaccurate names,
shadowed variables, overloaded temporaries (tmp, i, x reused across
unrelated blocks), and low-confidence blocks where names make it hard to
reason about the code.
Cohesion review looks for ten patterns of scattered code: distant configuration, split validation, fragmented type definitions, separated tests, dispersed error handling, remote utilities, disconnected docs, scattered state management, split domain logic, and orphaned dependencies.
Performance review analyzes five categories: algorithmic complexity, data structure efficiency, unnecessary work, I/O and external operations, and language-specific optimizations. Each finding includes expected improvement and tradeoffs.
Test coverage review checks for gaps in: happy paths, branch coverage, edge cases, error paths, boundary conditions, integration points, return values, state changes, and concurrency concerns.
Gridmat picks up to 5 entry points, assigns each a color-coded emoji, and traces execution paths downward through the codebase in an ASCII box-drawing diagram.
Running the server directly
The server uses stdio transport. To start it manually:
ctroy-code-analysis
Or:
python -m ctroy_code_analysis.server
