Latch - Security Guardrails for AI Agents

Open-source guard proxy for MCP servers. Stop your AI agent before it does something you’ll regret.

Introduction

AI agents are incredibly powerful. They can read your files, execute shell commands, send emails, query databases, and make API calls on your behalf.

But what happens when your agent decides to:

• Search for passwords in your home directory?
• Delete files it thinks are “temporary”?
• Send confidential data to the wrong recipient?
• Run a shell command that nukes your database?

You need a human in the loop. You need Latch.

What is Latch?

Latch is an open-source guard proxy that sits between your AI agent and its tools. Every tool call passes through Latch, which evaluates it against your security policies:

Action	What Happens
✅ Allowed operations	Proceed automatically when allowed by rule (or by default)
⚠️ Approval-required operations	Pause until you approve (when a rule requires approval)
🛑 Denied operations	Blocked entirely (when a rule denies)
🤖 LLM policies	Natural language conditions evaluated in real-time

Quick Example

# Without Latch
Agent: "Search for API keys in ~/"
→ 💥 Agent finds and potentially leaks your secrets

# With Latch
Agent: "Search for API keys in ~/"  
→ ⚠️ Requires approval (or 🛑 denied), depending on your rules
→ You see it in the audit log (with redacted arguments)
→ You decide what happens

Key Features

Policy Engine

Rules based on tool name, action class, domain, and recipient. Most-specific rule wins.

LLM-Evaluated Policies

Write policy conditions in plain English. An LLM evaluates each tool call against them.

Approval Workflow

Risky actions pause for your approval. Single-use tokens or time-limited leases.

Telegram Notifications

Get notified on your phone. Approve or deny with one tap.

Audit Log

Full history of every tool call, decision, and redacted arguments.

Self-Hosted

Run on your infrastructure. Your data never leaves your control.

Works With

Latch works with any MCP (Model Context Protocol) compatible client and server:

Clients:

• Claude Desktop
• Cursor
• Any MCP-compatible agent

Servers (examples):

• Filesystem access
• GitHub operations
• Database queries
• Shell commands
• Email/Slack
• Any custom MCP server

See It In Action

The Dashboard

Create workspaces, manage upstreams, and monitor everything in one place.

Policy Rules

Define what’s allowed, what’s blocked, and what needs approval.

Audit Log

Every request logged with the decision and redacted arguments.

Get Started in 5 Minutes

Quickstart

Run Latch locally with Docker and wrap your first MCP server.

Claude Desktop Setup

Step-by-step guide to integrate Latch with Claude Desktop.

Open Source

Latch is fully open source under the MIT license.

• GitHub Repository
• Report Issues
• Contributing Guide

---

Built for developers who want AI agents they can actually trust.