Skip to main content
Glama
llm-code-security-review-mcp

llm-code-security-review-mcp

by drewrukin

LLM Code Security Review MCP Server

This MCP server is a companion tool for the llm-code-security-review methodology. It helps an LLM follow the review plan one step at a time. The LLM performs the security review; the server tracks progress, makes sure the LLM actually reads all source code assigned to each subsystem, and does not accept the review as complete until every plan step is finished.

The Short Version

  1. Load a review plan.

  2. Codex receives one subsystem and one security area at a time.

  3. The server checks that the assigned source was read before accepting the result.

  4. Progress is saved automatically. You can restart Codex, pause the review, or continue later without starting over.

  5. Findings must be reviewed separately. The server organizes the work; it does not decide whether a reported issue is real.

That is the whole job.

Related MCP server: DevFlow MCP Server

What You Need

  • Codex on Linux, macOS, or another POSIX system;

  • Python 3.11 or newer;

  • a review plan created with the companion methodology.

The server has no third-party Python dependencies.

Install It

git clone https://github.com/drewrukin/llm-code-security-review-mcp.git
cd llm-code-security-review-mcp
python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install .

Add the server to ~/.codex/config.toml. Replace the example path with the absolute path to your clone:

[mcp_servers.llm-code-security-review-mcp]
command = "/absolute/path/to/llm-code-security-review-mcp/.venv/bin/llm-code-security-review-mcp-server"

Start a new Codex session after changing the configuration.

Run a Review

First, use the companion methodology to create a plan for your project. Then start a new Codex session and say:

Use the llm-code-security-review-mcp server to load <path-to-plan.json> and execute the review.

The server will provide the first task and keep the review moving until the plan is complete.

Development

Run the test suite with:

python3 -m unittest discover -s tests

Released changes are in CHANGELOG.md. Please report security problems privately as described in SECURITY.md.

License

Copyright 2026 Andrew Rukin (drewrukin). Licensed under the Apache License 2.0.

Available Tools

4 tools
get_taskA

Issue or repeat the one current security-review task. Obey its execution_instruction: read source serially in one-file or <=300-line calls, submit report_done for mechanical coverage verification, then immediately call get_task {} again.

Inputs: no named parameters.

load_planA

Load the supplied security-review plan before execution.

Inputs: path (required).

report_doneA

Submit the structured result. This call mechanically verifies source coverage and returns the exact missing read when incomplete. For NO FINDINGS, finding_ids, evidence, and artifacts must all be empty. For FINDING or CORROBORATION, include only concise source ranges needed to prove the security effect. After an accepted result, immediately call get_task {} and continue until COMPLETE; after COMPLETE, call status {} and use only its diagnostics for final totals.

Inputs: outcome (required), step_id (required), summary (required), evidence (required), +2 more.

statusA
Read-onlyIdempotent

Return read-only authoritative review progress and diagnostics. After the server reports COMPLETE, call status {} and use only its diagnostics for final totals.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a distinct role: load_plan initializes the review, get_task fetches the active task, status reports progress, and report_done submits results. Despite verbose instructions, the boundaries are clear and an agent should not confuse them.

    Naming Consistency4/5

    load_plan and get_task follow a clean verb_noun pattern, and all names use lowercase snake_case. status is a bare noun and report_done uses an adjective complement rather than a noun object, so the pattern is not perfect but remains readable and predictable.

    Tool Count5/5

    Four tools is well-scoped for a focused security-review orchestration server. Each tool maps to an essential phase of the workflow and none feels redundant or missing.

    Completeness5/5

    The tool surface covers the full intended lifecycle: load the plan, retrieve tasks, submit results, and check final status. There are no obvious dead ends or missing operations within the server's clearly defined orchestration purpose.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      D
      quality
      D
      maintenance
      An MCP server that implements a structured workflow for LLM-based coding, guiding development through feature clarification, documentation generation, phased implementation, and progress tracking.
      10
      58 npm
      17
      MIT
    • A
      license
      B
      quality
      B
      maintenance
      MCP server that integrates DevFlow with AI code assistants to enforce structured development workflows including planning, task tracking, and code review gates.
      65
      15 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server that securely reviews AI-generated code for vulnerabilities such as SQL injection, command injection, and hardcoded credentials.
      12
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      A lightweight MCP server for security reviews that injects security requirements before code generation, scans dependencies for CVEs, and verifies generated code without disrupting workflow.
      73
      MIT