Code Firewall MCP

Code Firewall MCP

A structural similarity-based code security filter for MCP (Model Context Protocol). Blocks dangerous code patterns before they reach execution tools by comparing code structure against a blacklist of known-bad patterns.

How It Works

flowchart LR

A[Code file/string] --> B[Parse & Normalize tree-sitter] B --> C[Embed Ollama] C --> D{Similarity Check vs Blacklist} D -->|≥ threshold| E[🚫 BLOCKED] D -->|< threshold| F[✅ ALLOWED] F --> G[Execution Tools rlm_exec, etc.]

style E fill:#ff6b6b,color:#fff style F fill:#51cf66,color:#fff style D fill:#339af0,color:#fff

1. Parse code to Concrete Syntax Tree (CST) using tree-sitter
2. Normalize by stripping identifiers and literals → structural skeleton
3. Embed the normalized structure via Ollama
4. Compare against blacklisted patterns in ChromaDB
5. Block if similarity exceeds threshold, otherwise allow

Key Insight

Code patterns like os.system("rm -rf /") and os.system("ls") have identical structure. By normalizing away the specific commands/identifiers, we can detect dangerous patterns regardless of the specific arguments used.

Security-sensitive identifiers are preserved during normalization (e.g., eval, exec, os, system, subprocess, Popen, shell) to ensure embeddings remain discriminative for dangerous patterns.

Installation

Quick Start

Option 1: PyPI (Recommended)

uvx code-firewall-mcp

or

pip install code-firewall-mcp

Option 2: Claude Desktop One-Click

Download the .mcpb from Releases and double-click to install.

Option 3: From Source

git clone https://github.com/egoughnour/code-firewall-mcp.git

cd code-firewall-mcp uv sync

Wire to Claude Code / Claude Desktop

Add to ~/.claude/.mcp.json (Claude Code) or claude_desktop_config.json (Claude Desktop):

{

"mcpServers": { "code-firewall": { "command": "uvx", "args": ["code-firewall-mcp"], "env": { "FIREWALL_DATA_DIR": "~/.code-firewall", "OLLAMA_URL": "http://localhost:11434" } } } }

Requirements

• Python 3.10+ (< 3.14 due to onnxruntime compatibility)
• Ollama (for embeddings)
• ChromaDB (for vector storage)
• tree-sitter (optional, for better parsing)

Setting Up Ollama (Embeddings)

Code Firewall can automatically install and configure Ollama on macOS with Apple Silicon. There are two installation methods:

Method 1: Homebrew Installation

# 1. Check system requirements

firewall_system_check()

2. Install via Homebrew

firewall_setup_ollama(install=True, start_service=True, pull_model=True)

What this does:

• Installs Ollama via Homebrew (brew install ollama)
• Starts Ollama as a managed background service
• Pulls nomic-embed-text model for embeddings

Method 2: Direct Download (No Sudo)

# 1. Check system

firewall_system_check()

2. Install via direct download - no sudo, no Homebrew

firewall_setup_ollama_direct(install=True, start_service=True, pull_model=True)

What this does:

• Downloads Ollama from https://ollama.com
• Extracts to ~/Applications/ (no admin needed)
• Starts Ollama via ollama serve
• Pulls nomic-embed-text model

Manual Setup

# Install Ollama

brew install ollama

or download from https://ollama.ai

Start service

brew services start ollama

or: ollama serve

Pull embedding model

ollama pull nomic-embed-text

Verify

firewall_ollama_status()

Tools

Setup & Status Tools

Firewall Tools

firewall_check

Check if a code file is safe to pass to execution tools.

result = await firewall_check(file_path="/path/to/script.py")

Returns: {allowed: bool, blocked: bool, similarity: float, ...}

firewall_check_code

Check code string directly (no file required).

result = await firewall_check_code(

code="import os; os.system('rm -rf /')", language="python" )

firewall_blacklist

Add a dangerous pattern to the blacklist.

result = await firewall_blacklist(

code="os.system(arbitrary_command)", reason="Arbitrary command execution", severity="critical" )

firewall_record_delta

Record near-miss variants to sharpen the classifier.

result = await firewall_record_delta(

code="subprocess.run(['ls', '-la'])", similar_to="abc123", notes="Legitimate use case for file listing" )

firewall_list_patterns

List patterns in the blacklist or delta collection.

firewall_remove_pattern

Remove a pattern from blacklist or deltas.

firewall_status

Get firewall status and statistics.

Configuration

Environment variables:

Usage Pattern

Pre-filter for massive-context-mcp

Use code-firewall-mcp as a gatekeeper before passing code to rlm_exec:

# 1. Check code safety

check = await firewall_check_code(user_code)

if check["blocked"]: print(f"BLOCKED: {check['reason']}") return

2. If allowed, proceed with execution

result = await rlm_exec(code=user_code, context_name="my-context")

Integrated with massive-context-mcp

Install massive-context-mcp with firewall integration:

pip install massive-context-mcp[firewall]

When enabled, rlm_exec automatically checks code against the firewall before execution.

Building the Blacklist

The blacklist grows through use:

1. Initial seeding: Add known dangerous patterns
2. Audit feedback: When rlm_auto_analyze finds security issues, add patterns
3. Delta sharpening: Record near-misses to improve classification boundaries

# After security audit finds issues

await firewall_blacklist( code=dangerous_code, reason="Command injection via subprocess", severity="critical" )

Structural Normalization

flowchart TD

subgraph Input A1["os.system('rm -rf /')"] A2["os.system('ls -la')"] A3["os.system(user_cmd)"] end

subgraph Normalization B[Strip literals & identifiers Preserve security keywords] end

subgraph Output C["os.system('S')"] end

A1 --> B A2 --> B A3 --> B B --> C

style C fill:#ff922b,color:#fff

The normalizer strips:

• Identifiers: my_var → _ (except security-sensitive ones)
• String literals: "hello" → "S"
• Numbers: 42 → N
• Comments: Removed entirely

Preserved identifiers (for better pattern matching):

• eval, exec, compile, __import__
• os, system, popen, subprocess, Popen, shell
• open, read, write, socket, connect
• getattr, setattr, __globals__, __builtins__
• And more security-sensitive names...

Example:

# Original

subprocess.run(["curl", url, "-o", output_file])

Normalized (preserves 'subprocess' and 'run')

subprocess.run(["S", _, "S", _])

Both subprocess.run(["curl", ...]) and subprocess.run(["wget", ...]) normalize to the same structure, so blacklisting one catches both.

License

MIT