47 lines
2.9 KiB
Markdown
47 lines
2.9 KiB
Markdown
|
# Product Design: The MCP Constraint Builder (Standalone)
|
||
|
|
|
||
|
|
This document outlines the product design and system architecture for a standalone security product: **The MCP Constraint Builder**. This tool acts as an auditing and policy-enforcement layer for AI agents, allowing users to visually configure execution policies and parameters for third-party MCP servers.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 1. Core Concept
|
||
|
|
|
||
|
|
The MCP Constraint Builder is a **security firewall and configuration manager** for local AI agents. It intercepts JSON-RPC requests sent by AI clients (such as Claude, Cursor, or local coding assistants) to local MCP servers.
|
||
|
|
|
||
|
|
```
|
||
|
|
+------------+ JSON-RPC +------------------------+ JSON-RPC +------------------+
|
||
|
|
| AI Client | -----------------> | MCP Policy Proxy | -----------------> | Target MCP Server|
|
||
|
|
| (Claude/etc| | (Enforces rules.json) | | (Proprietary) |
|
||
|
|
+------------+ +-----------+------------+ +------------------+
|
||
|
|
|
|
||
|
|
| Reads policies
|
||
|
|
v
|
||
|
|
+------------------------+
|
||
|
|
| rules.json |
|
||
|
|
+------------------------+
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 2. Key Features
|
||
|
|
|
||
|
|
### Visual Constraint Editor
|
||
|
|
A desktop application that allows users to import any standard MCP server configuration. The tool inspects the server's available tools (`list_tools`) and dynamically renders a visual configuration interface:
|
||
|
|
* **Allow/Deny lists**: Explicitly toggle which tools (e.g. `run_command`, `read_file`) the agent is allowed to access.
|
||
|
|
* **Parameter boundaries**: Set regex patterns or value ranges for arguments (e.g. "Only allow `read_file` if the path starts with `C:/Users/Owner/workspace`").
|
||
|
|
* **Execution budget**: Set rate limits (e.g. "maximum 10 tool calls per minute").
|
||
|
|
|
||
|
|
### Proxy Firewall (JSON-RPC Middleware)
|
||
|
|
A lightweight background proxy runner. When the AI agent attempts to execute a tool, the proxy checks the arguments against `rules.json`:
|
||
|
|
* **Pass**: If within bounds, forwards the request directly to the target server.
|
||
|
|
* **Block**: If out of bounds, immediately returns an authorization error back to the AI client without executing the command.
|
||
|
|
* **Escalate**: If marked for manual verification, prompts the user on their host machine for real-time bypass approval.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 3. Benefits
|
||
|
|
|
||
|
|
1. **Enterprise Compliance**: Solves the data protection and code execution trust gap for running local agents.
|
||
|
|
2. **IP Protection**: Allows creators of advanced AI agents to sell compiled/obfuscated MCP binaries while giving users full visibility and boundary control.
|
||
|
|
3. **Ecosystem Agnostic**: Works with any standard MCP client and server protocol without requiring code changes to either.
|