mcpwall

iptables for MCP. Blocks dangerous tool calls, scans for secret leakage, logs everything. No AI, no cloud, pure rules.

Sits between your AI coding tool (Claude Code, Cursor, Windsurf) and MCP servers, intercepting every JSON-RPC message and enforcing YAML-defined policies.

Why

MCP servers have full access to your filesystem, shell, databases, and APIs. When an AI agent calls tools/call, the server executes whatever the agent asks — reading SSH keys, running rm -rf, exfiltrating secrets. There's no built-in policy layer.

mcpwall adds one. It's a transparent stdio proxy that:

• Blocks sensitive file access — .ssh/, .env, credentials, browser data

• Blocks dangerous commands — rm -rf, pipe-to-shell, reverse shells

• Scans for secret leakage — API keys, tokens, private keys (regex + entropy)

• Scans server responses — redacts leaked secrets, blocks prompt injection patterns, flags suspicious content

• Logs everything — JSON Lines audit trail of every tool call and response

• Uses zero AI — deterministic rules, no LLM decisions, no cloud calls

Install

Or use directly with npx:

Quick Start

Option 1: Docker MCP Toolkit

If you use Docker MCP Toolkit (the most common setup), change your MCP config from:

To:

That's it. mcpwall now sits in front of all your Docker MCP servers, logging every tool call and blocking dangerous ones. No config file needed — sensible defaults apply automatically.

Option 2: Interactive setup

This finds your existing MCP servers in Claude Code, Cursor, Windsurf, and VS Code configs and wraps them.

Option 3: Manual wrapping (any MCP server)

Change your MCP config from:

To:

Option 4: Wrap a specific server

How It Works

Inbound (requests):

1. Intercepts every JSON-RPC request on stdin

2. Parses tools/call requests — extracts tool name and arguments

3. Walks rules top-to-bottom, first match wins

4. Allow: forward to real server

5. Deny: return JSON-RPC error to host, log, do not forward

Outbound (responses):

1. Parses every response from the server before forwarding

2. Evaluates against outbound_rules (same first-match-wins semantics)

3. Allow: forward unchanged

4. Deny: replace response with blocked message

5. Redact: surgically replace secrets with [REDACTED BY MCPWALL], forward modified response

6. Log only: forward unchanged, log the match

Configuration

Config is YAML. mcpwall looks for:

1. ~/.mcpwall/config.yml (global)

2. .mcpwall.yml (project, overrides global)

If neither exists, built-in default rules apply.

Example config

Rule matchers

The special key _any_value applies the matcher to ALL argument values.

Outbound rules (response inspection)

Outbound rules scan server responses before they reach your AI client. Add them to the same config file:

Outbound matchers

Outbound actions

Built-in rule packs

• rules/default.yml — sensible defaults (blocks SSH, .env, credentials, dangerous commands, secrets)

• rules/strict.yml — deny-by-default paranoid mode (whitelist only project reads/writes)

Use strict mode:

CLI

Options:

• -c, --config <path> — path to config file

• --log-level <level> — override log level (debug/info/warn/error)

Audit Logs

All tool calls are logged by default — both allowed and denied. Logs are written as JSON Lines to ~/.mcpwall/logs/YYYY-MM-DD.jsonl:

Denied entries have args redacted to prevent secrets from leaking into logs.

mcpwall also prints color-coded output to stderr so you can see decisions in real time.

Security Design

• Bidirectional scanning: Both inbound requests and outbound responses are evaluated against rules

• Fail closed on invalid config: Bad regex in a rule crashes at startup, never silently passes traffic

• Fail open on outbound errors: If response parsing fails, the raw response is forwarded (never blocks legitimate traffic)

• Args redacted on deny: Blocked tool call arguments are never written to logs

• Surgical redaction: Secrets in responses are replaced in-place, preserving the JSON-RPC response structure

• Path traversal defense: not_under matcher uses path.resolve() to prevent ../ bypass

• Pre-compiled regexes: All patterns compiled once at startup for consistent performance

• No network: Zero cloud calls, zero telemetry, runs entirely local

• Deterministic: Same input + same rules = same output, every time

License

Apache-2.0

mcpwall is not affiliated with or endorsed by Anthropic or the Model Context Protocol project. MCP is an open protocol maintained by the Agentic AI Foundation under the Linux Foundation.